+ All Categories
Home > Documents > Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf ·...

Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf ·...

Date post: 03-Aug-2020
Category:
Upload: others
View: 3 times
Download: 0 times
Share this document with a friend
42
Dynamic reconfiguration of human brain networks during learning Danielle S. Bassett a,1 , Nicholas F. Wymbs b , Mason A. Porter c,d , Peter J. Mucha e,f , Jean M. Carlson a , and Scott T. Grafton b a Complex Systems Group, Department of Physics, University of California, Santa Barbara, CA 93106; b Department of Psychology and UCSB Brain Imaging Center, University of California, Santa Barbara, CA 93106; c Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute, University of Oxford, Oxford OX1 3LB, United Kingdom; d Complex Agent-Based Dynamic Networks Complexity Centre, University of Oxford, Oxford OX1 1HP, United Kingdom; e Carolina Center for Interdisciplinary Applied Mathematics, Department of Mathematics, University of North Carolina, Chapel Hill, NC 27599; and f Institute for Advanced Materials, Nanoscience and Technology, University of North Carolina, Chapel Hill, NC 27599 Edited by Marcus E. Raichle, Washington University in St. Louis, St. Louis, MO, and approved March 15, 2011 (received for review December 16, 2010) Human learning is a complex phenomenon requiring flexibility to adapt existing brain function and precision in selecting new neu- rophysiological activities to drive desired behavior. These two attributesflexibility and selectionmust operate over multiple temporal scales as performance of a skill changes from being slow and challenging to being fast and automatic. Such selective adapt- ability is naturally provided by modular structure, which plays a critical role in evolution, development, and optimal network func- tion. Using functional connectivity measurements of brain activity acquired from initial training through mastery of a simple motor skill, we investigate the role of modularity in human learning by identifying dynamic changes of modular organization spanning multiple temporal scales. Our results indicate that flexibility, which we measure by the allegiance of nodes to modules, in one experi- mental session predicts the relative amount of learning in a future session. We also develop a general statistical framework for the identification of modular architectures in evolving systems, which is broadly applicable to disciplines where network adaptability is crucial to the understanding of system performance. complex network time-dependent network fMRI motor learning community structure T he brain is a complex system, composed of many interacting parts, which dynamically adapts to a continually changing environment over multiple temporal scales. Over relatively short temporal scales, rapid adaptation and continuous evolution of those interactions or connections form the neurophysiological basis for behavioral adaptation or learning. At small spatial scales, stable neurophysiological signatures of learning have been best demonstrated in animal systems at the level of individual synapses between neurons (13). At a larger spatial scale, it is also well-known that specific regional changes in brain activity and effective connectivity accompany many forms of learning in humansincluding the acquisition of motor skills (4, 5). Learning-associated adaptability is thought to stem from the principle of cortical modularity (6). Modular, or nearly decom- posable (7), structures are aggregates of small subsystems (mod- ules) that can perform specific functions without perturbing the remainder of the system. Such structure provides a combination of compartmentalization and redundancy, which reduces the interdependence of components, enhances robustness, and facil- itates behavioral adaptation (8, 9). Modular organization also confers evolvability on a system by reducing constraints on change (8, 1012). Indeed, a putative relationship between mod- ularity and adaptability in the context of human neuroscience has recently been posited (13, 14). To date, however, the existence of modularity in large-scale cortical connectivity during learning has not been tested directly. Based on the aforementioned theoretical and empirical grounds, we hypothesized that the principle of modularity would characterize the fundamental organization of human brain func- tional connectivity during learning. More specifically, based on several studies relating the neural basis of modularity to the development of skilled movements (1517), we expected that functional brain networks derived from acquisition of a simple motor skill would display modular structure over the variety of temporal scales associated with learning (18). We also hypothe- sized that modular structure would change dynamically during learning (4, 19), and that characteristics of such dynamics would be associated with learning success. We tested these predictions using fMRI, an indirect measure of local neuronal activity (20), in healthy adult subjects during the acquisition of a simple motor learning skill composed of visually cued finger sequences. We derived low frequency (0.060.12 Hz) functional networks from the fMRI data by computing the tem- poral correlation between activity in each pair of brain regions to construct weighted graphs or whole-brain functional networks (2123) (Fig. 1A and SI Appendix). This network framework enabled us to estimate a mathematical representation of modular or community organization, known as network modularity,for each individual over a range of temporal scales. We evaluated the evolution of network connectivity over time using the mathema- tical framework described in ref. 25, and we tested its relationship with learning. See Materials and Methods for details of the sample, experimental paradigm, and methods of analysis. Results Static Modular Structure. We investigated network organization over multiple temporal scalesover days, hours, and minutes during motor learning (18, 19) (Fig. 1B). We used a diagnostic measure of the amount of network modularity in the systemthe modularity index Q (See Materials and Methods for a mathema- tical definition). At each scale, we found Q to be larger than ex- pected in a random network, indicating a significant segregation of the brain into distinct modules or communities (Fig. 2 AC). The cortex is organized into fewer modules than the random network, indicating that the functional activity of the brain is sig- nificantly integrated across cortical regions. Because these results were consistent for all of the temporal scales that we examined, we concluded that the brain shows temporal scaling of functional organization, consistent with the scaling in frequency (26) and spatial (27, 28) domains previously reported. Furthermore, the temporal structure of this organization is graded in the sense that fewer modules (about three) on longer timescales (Fig. 2 A and B) are complemented by more modules (about four) on shorter timescales (Fig. 2C). This graded structure is analogous to that found in the nested modular networks of underlying brain Author contributions: D.S.B., N.F.W., M.A.P., P.J.M., and S.T.G. designed research; D.S.B. and N.F.W. performed research; D.S.B., N.F.W., M.A.P., P.J.M., J.M.C., and S.T.G. contributed new reagents/analytic tools; D.S.B. and P.J.M. wrote the code; D.S.B. analyzed data; and D. S.B., N.F.W., and M.A.P. wrote the paper. The authors declare no conflict of interest. This article is a PNAS Direct Submission. 1 To whom correspondence should be addressed. E-mail: [email protected]. This article contains supporting information online at www.pnas.org/lookup/suppl/ doi:10.1073/pnas.1018985108/-/DCSupplemental. www.pnas.org/cgi/doi/10.1073/pnas.1018985108 PNAS May 3, 2011 vol. 108 no. 18 76417646 SYSTEMS BIOLOGY APPLIED MATHEMATICS
Transcript
Page 1: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Dynamic reconfiguration of human brainnetworks during learningDanielle S. Bassetta,1, Nicholas F. Wymbsb, Mason A. Porterc,d, Peter J. Muchae,f, Jean M. Carlsona, and Scott T. Graftonb

aComplex Systems Group, Department of Physics, University of California, Santa Barbara, CA 93106; bDepartment of Psychology and UCSB Brain ImagingCenter, University of California, Santa Barbara, CA 93106; cOxford Centre for Industrial and Applied Mathematics, Mathematical Institute, University ofOxford, Oxford OX1 3LB, United Kingdom; dComplex Agent-Based Dynamic Networks Complexity Centre, University of Oxford, Oxford OX1 1HP, UnitedKingdom; eCarolina Center for Interdisciplinary Applied Mathematics, Department of Mathematics, University of North Carolina, Chapel Hill, NC 27599;and fInstitute for Advanced Materials, Nanoscience and Technology, University of North Carolina, Chapel Hill, NC 27599

Edited by Marcus E. Raichle, Washington University in St. Louis, St. Louis, MO, and approved March 15, 2011 (received for review December 16, 2010)

Human learning is a complex phenomenon requiring flexibility toadapt existing brain function and precision in selecting new neu-rophysiological activities to drive desired behavior. These twoattributes—flexibility and selection—must operate over multipletemporal scales as performance of a skill changes from being slowand challenging to being fast and automatic. Such selective adapt-ability is naturally provided by modular structure, which plays acritical role in evolution, development, and optimal network func-tion. Using functional connectivity measurements of brain activityacquired from initial training through mastery of a simple motorskill, we investigate the role of modularity in human learning byidentifying dynamic changes of modular organization spanningmultiple temporal scales. Our results indicate that flexibility, whichwe measure by the allegiance of nodes to modules, in one experi-mental session predicts the relative amount of learning in a futuresession. We also develop a general statistical framework for theidentification of modular architectures in evolving systems, whichis broadly applicable to disciplines where network adaptability iscrucial to the understanding of system performance.

complex network ! time-dependent network ! fMRI ! motor learning !community structure

The brain is a complex system, composed of many interactingparts, which dynamically adapts to a continually changing

environment over multiple temporal scales. Over relatively shorttemporal scales, rapid adaptation and continuous evolution ofthose interactions or connections form the neurophysiologicalbasis for behavioral adaptation or learning. At small spatialscales, stable neurophysiological signatures of learning have beenbest demonstrated in animal systems at the level of individualsynapses between neurons (1–3). At a larger spatial scale, it isalso well-known that specific regional changes in brain activityand effective connectivity accompany many forms of learningin humans—including the acquisition of motor skills (4, 5).

Learning-associated adaptability is thought to stem from theprinciple of cortical modularity (6). Modular, or nearly decom-posable (7), structures are aggregates of small subsystems (mod-ules) that can perform specific functions without perturbing theremainder of the system. Such structure provides a combinationof compartmentalization and redundancy, which reduces theinterdependence of components, enhances robustness, and facil-itates behavioral adaptation (8, 9). Modular organization alsoconfers evolvability on a system by reducing constraints onchange (8, 10–12). Indeed, a putative relationship between mod-ularity and adaptability in the context of human neuroscience hasrecently been posited (13, 14). To date, however, the existence ofmodularity in large-scale cortical connectivity during learning hasnot been tested directly.

Based on the aforementioned theoretical and empiricalgrounds, we hypothesized that the principle of modularity wouldcharacterize the fundamental organization of human brain func-tional connectivity during learning. More specifically, based onseveral studies relating the neural basis of modularity to the

development of skilled movements (15–17), we expected thatfunctional brain networks derived from acquisition of a simplemotor skill would display modular structure over the variety oftemporal scales associated with learning (18). We also hypothe-sized that modular structure would change dynamically duringlearning (4, 19), and that characteristics of such dynamics wouldbe associated with learning success.

We tested these predictions using fMRI, an indirect measureof local neuronal activity (20), in healthy adult subjects during theacquisition of a simple motor learning skill composed of visuallycued finger sequences. We derived low frequency (0.06–0.12 Hz)functional networks from the fMRI data by computing the tem-poral correlation between activity in each pair of brain regions toconstruct weighted graphs or whole-brain functional networks(21–23) (Fig. 1A and SI Appendix). This network frameworkenabled us to estimate a mathematical representation of modularor community organization, known as “network modularity,” foreach individual over a range of temporal scales. We evaluated theevolution of network connectivity over time using the mathema-tical framework described in ref. 25, and we tested its relationshipwith learning. SeeMaterials andMethods for details of the sample,experimental paradigm, and methods of analysis.

ResultsStatic Modular Structure. We investigated network organizationover multiple temporal scales—over days, hours, and minutes—during motor learning (18, 19) (Fig. 1B). We used a diagnosticmeasure of the amount of network modularity in the system—themodularity index Q (See Materials and Methods for a mathema-tical definition). At each scale, we found Q to be larger than ex-pected in a random network, indicating a significant segregationof the brain into distinct modules or communities (Fig. 2 A–C).The cortex is organized into fewer modules than the randomnetwork, indicating that the functional activity of the brain is sig-nificantly integrated across cortical regions. Because these resultswere consistent for all of the temporal scales that we examined,we concluded that the brain shows temporal scaling of functionalorganization, consistent with the scaling in frequency (26) andspatial (27, 28) domains previously reported. Furthermore, thetemporal structure of this organization is graded in the sensethat fewer modules (about three) on longer timescales (Fig. 2 Aand B) are complemented by more modules (about four) onshorter timescales (Fig. 2C). This graded structure is analogousto that found in the nested modular networks of underlying brain

Author contributions: D.S.B., N.F.W., M.A.P., P.J.M., and S.T.G. designed research; D.S.B.and N.F.W. performed research; D.S.B., N.F.W., M.A.P., P.J.M., J.M.C., and S.T.G. contributednew reagents/analytic tools; D.S.B. and P.J.M. wrote the code; D.S.B. analyzed data; and D.S.B., N.F.W., and M.A.P. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.1To whom correspondence should be addressed. E-mail: [email protected].

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1018985108/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1018985108 PNAS ! May 3, 2011 ! vol. 108 ! no. 18 ! 7641–7646

SYST

EMSBIOLO

GY

APP

LIED

MAT

HEMAT

ICS

Page 2: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

anatomy where few modules uncovered at large spatial scales arecomplemented by more modules at smaller spatial scales (27).

Dynamic Modular Structure.We next consider evolvability, which ismost readily detected when the organism is under stress (29) orwhen acquiring new capacities such as during external training inour experiment. We found that the community organization ofbrain connectivity reconfigured adaptively over time. Using a re-cently developed mathematical formalism to assess the presenceof dynamic network reconfigurations (25), we constructed multi-layer networks in which we link the network for each time window(Fig. 3A) to the network in the time windows before and after(Fig. 3B) by connecting each node to itself in the neighboring win-dows. We then measured modular organization (30–32) on thislinked multilayered network to find long-lasting modules (25).

To verify the reliability of our measurements of dynamic mod-ular architecture, we introduced three null models based on per-mutation testing (Fig. 3C). We found that cortical connectivity isspecifically patterned, which we concluded by comparison to a“connectional” null model in which we scrambled links betweennodes in each time window (33). Furthermore, cortical regionsmaintain these individual connectivity signatures that definecommunity organization, which we concluded by comparison toa “nodal” null model in which we linked a node in one time win-dow to a randomly chosen node in the previous and next timewindows. Finally, we found that functional communities exhibita smooth temporal evolution, which we identified by comparingdiagnostics computed using the true multilayer network structureto those computed using a temporally permuted version (Fig. 3D).We constructed this temporal null model by randomly reorderingthe multilayer network layers in time.

By comparing the structure of the cortical network to thoseof the null models, we found that the human brain exhibited aheightened modular structure in which more modules of smallersize were discriminable as a consequence of the emergence andextinction of modules in cortical network evolution. The statio-narity of communities, defined by the average correlation be-tween partitions over consecutive time steps (34), was also higherin the human brain than in the connectional or nodal null models,indicating a smooth temporal evolution.

Learning. Given the dynamic architecture of brain connectivity, itis interesting to ask whether the specific architecture changes

A

B

Fig. 1. Structure of the investigation. (A) To characterize the network struc-ture of low-frequency functional connectivity (24) at each temporal scale,we partitioned the raw fMRI data (Upper Left) from each subject’s brain intosignals originating from N ! 112 cortical structures, which constitute the net-work’s nodes (Upper Right). The functional connectivity, constituting the net-work edges, between two cortical structures is given by a Pearson correlationbetween the mean regional activity signals (Lower Right). We then statisti-cally corrected the resulting N ! N correlation matrix using a false discoveryrate correction (54) to construct a subject-specific weighted functional brainnetwork (Lower Left). (B) Schematic of the investigation that was performedover the temporal scales of days, hours, and minutes. The complete experi-ment, which defines the largest scale, took place over the course of threedays. At the intermediate scale, we conducted further investigations ofthe experimental sessions that occurred on each of those three days. Finally,to examine higher-frequency temporal structure, we cut each experimentalsession into 25 nonoverlapping windows, each of which was a fewminutes induration.

A C

B

Fig. 2. Multiscale modular architecture. (A) Results for the modular decomposition of functional connectivity across temporal scales. (Left) The network plotsshow the extracted modules; different colors indicate different modules and larger separation between modules is used to visualize weaker connectionsbetween them. (A) and (B) correspond to the entire experiment and individual sessions, respectively. Boxplots show the modularity index Q (Left)and the number of modules (Right) in the brain network compared to randomized networks. See Materials and Methods for a formal definition of Q.(C) Modularity index Q and the number of modules for the cortical (blue) compared to randomized networks (red) over the 75 time windows. Error barsindicate standard deviation in the mean over subjects.

7642 ! www.pnas.org/cgi/doi/10.1073/pnas.1018985108 Bassett et al.

Page 3: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

with learning—either at a gross scale through an adaptation inthe number or sizes of modules or at a finer scale through altera-tions in the nodal composition of modules. Empirically, we foundno significant differences between experimental sessions in thecoarse diagnostics. To quantify finer-scale architectural fluctua-tions, we introduced the notion of node flexibility using thenetwork properties determined in the multilayer framework.“Flexibility” is the number of times that each node changesmodule allegiance, normalized by the total possible number ofchanges (SI Appendix). The flexibility of the network as a wholeis then defined as the mean flexibility over all nodes.

Network flexibility is a measure that captures changes in thelocal properties of individual network elements. We found thatnetwork flexibility changed during the learning process—firstincreasing and then decreasing (Fig. 4A)—demonstrating ameaningful biological process. In particular, the flexibility of aparticipant in one session could be used as a predictor of theamount of learning (as measured by improvement in the timerequired to complete the sequence of motor responses) in thefollowing session (Fig. 4B). Regions of the brain that were mostresponsible for this predictive power of individual differences in

learning were distributed throughout the cortex, with strong load-ings in the frontal, presupplementary motor, posterior parietal,and occipital cortices (Fig. 4 C and D). We could not predictfuture learning capacity reliably using conventional task-relatedfMRI activation, supporting our conclusion that flexibility pro-vides a useful approach for modeling system evolvability.

Our results indicate that flexibility is sensitive to both intra-individual and interindividual variability. Across participants,we found that network flexibility was modulated by learning(Fig. 4A). However, we also found that each participant displayeda characteristic flexibility. The variation in flexibility over parti-cipants was larger than the variation in flexibility across sessions,as measured by the intraclass correlation coefficient: ICC " 0.56,F-statistic F"17;34# " 4.85, p " 4 ! 10#5.

DiscussionModularity of Functional Connectivity. Modularity is an intuitivelyimportant property for dynamic, adaptable systems. The accom-panying system decomposability provides necessary structure forcomplex reconfigurations. Modularity can be a property of mor-phology, as has been widely described in the context of evolutionand development (11, 12, 29), as well as of the interconnectionpatterns of social, biological, and technological systems (30, 31).More pertinent to this paper, recent evidence suggests that mod-ular organization over several spatial scales, or hierarchical mod-ularity, also characterizes the large-scale anatomical connectivityof the human brain (27, 28), as well as the spontaneous fluctua-tions (35, 36) thought to stem from anatomical patterns (37).However, the putative relationship between adaptability andmodular structure has not been previously explored in the contextof the brain connectome.

In the present study, we have shown that the functional con-nectivity of the human brain during a simple learning paradigmis inhomogeneous. Instead, it is segregated into communities thatcan each perform unique functions. This segregation of connec-tivity structure manifested consistently over the scale of days,hours, and minutes, suggesting that community structure providesa generalizable framework to study the evolution of temporallydistinct phenomena (12). However, it is also notable that connec-tivity at the shortest temporal scale displayed higher variability,perhaps reflecting the necessity for dynamic modulation of hu-man brain function over relatively short intervals during learning(19). In light of historically strict definitions of cognitive modulesas completely encapsulated structures (38), it is important toemphasize that the modules that we have uncovered remain in-tegrated with one another by a complex pattern of weak intercon-nections.

Dynamic Network Evolution. Efforts to characterize both restingstate (39) and task-based large-scale connectivity of human brainstructure and function (21–23) have focused almost exclusivelyon static representations of underlying connectivity patterns.However, both scientific intuition and recent evidence suggestthat connectivity can be modulated both spontaneously (40) andby exogenous stimulation (4). The exploration of temporally evol-ving network architecture therefore forms a critical frontier inneuroscience.

Our exploration of dynamic community structure in an experi-mental paradigm that requires neurophysiological adaptabilityprovides insight into the organizational principles supporting suc-cessful brain dynamics. Similar to social systems (34), we foundthat community organization changed smoothly with time, dis-playing coherent temporal dependence on what had gone beforeand what came after, a characteristic compatible with complexlong-memory dynamical systems (41).

In addition to global adaptability, we found that diverse re-gions of the brain performed different roles within communities:Some maintain community allegiance throughout the experiment

Fig. 3. Temporal dynamics of modular architecture. (A) Schematic of a toynetwork with four nodes and four edges in a single time window. (B) Multi-layer network framework in which the networks from four time windows arelinked by connecting nodes in one time window to themselves in the adja-cent time windows (colored curves). (C) Statistical framework composed of aconnectional null model (Top), a nodal null model (Middle), and a temporalnull model (Bottom) in which intranetwork links, internetwork links, andtime windows, respectively, in the real network are randomized in the per-muted network. (We show all of the randomized links in red.) (D) Boxplotsshowing differences in modular architecture between the real and permutednetworks for the connectional (Top), nodal (Middle), and temporal (Bottom)null models. We measured the structure of the network using the modularityindex Q, the number of modules, the module size, and stationarity, which isdefined as the mean similarity in the nodal composition of modules over con-secutive time steps. Below each plot, we indicate by asterisks the significanceof one-sample t-tests that assess whether the differences that we observedwere significantly different from zero (gray lines): A single asterisk indicatesp < :05, two asterisks indicate p < 1 ! 10#6, and three asterisks indicatep < 1 ! 10#20.

Bassett et al. PNAS ! May 3, 2011 ! vol. 108 ! no. 18 ! 7643

SYST

EMSBIOLO

GY

APP

LIED

MAT

HEMAT

ICS

Page 4: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

(low-flexibility nodes), and others constantly shift allegiance(high-flexibility nodes). Biologically, this network flexibility mightbe driven by physiological processes that facilitate the participa-tion of cortical regions in multiple functional communities.Learning a motor skill induces changes in both the structure andconnectivity of the cortex (42, 43), which is accompanied by in-creased excitability and decreased inhibition of neural circuitry(44–46). However, it is plausible that flexibility might also bedriven by task-dependent processes that require the capacity tobalance learning across subtasks. For example, the particular ex-periment utilized in this study demanded that subjects master theuse of a response box, decoding of the stimulus, performance ofprecise movements, balancing of attention between stimuli, andswitching between different sequences of movements.

Flexibility and Learning. Importantly, the inherent temporal varia-bility in network structure measured by nodal flexibility was not astable signature of an individual’s functional organization butwas instead modulated by consecutive stages of learning—firstincreasing and then decreasing as movement time stabilized inthe later stages of learning (19). The modulation of flexibilityby learning was evident not only at the group level but also inindividuals. The amount of flexibility in each participant couldbe used to predict that participant’s learning in a following experi-mental session. In addition to supporting the theoretical utility ofaccessible but often ignored higher-order (bivariate, multivariate)statistics of brain function, this result could potentially be used toinform decisions on how and when to train individuals on newtasks depending on the current flexibility of their brain. From thiswork alone, however, we are unable to determine whether or notlearning is the only possible modulator of flexibility. Complemen-tary experiments could be designed to test whether flexibility isalso modulated by fatigue or exogenous stimulants to increase

subsequent skilled learning. We also found that interindividualvariability in flexibility was larger than intraindividual variability,indicating that flexibility might be a reliable indicator of a givensubject’s brain state. Consequently, our methodology couldpotentially be of use in predicting a given individual’s responseto training or neurorehabilitation (47, 48).

Flexibility might be a network signature of a complex under-lying cortical system characterized by noise (49). Such a hypo-thesis is bolstered by recent complementary evidence suggestingthat variability in brain signals also supports mental effort in avariety of cognitive operations (50), presumably by aiding thebrain in switching between different network configurations asit masters a new task. Indeed, the theoretical utility of noise ina nonlinear dynamical system like the brain (51) lies in its facil-itation of transitions between network states or system functions(52) and therefore helps to delineate the system’s dynamic reper-toire (53). However, despite the plausibility that network flexibil-ity and cortical noise are related, future studies are necessary todirectly test this hypothesis.

Methodological Considerations.The construction of brain networksfrom continuous association matrices, such as those based onpairwise correlation or coherence, has historically been per-formed by applying a threshold to the data to construct a binarygraph in which an edge exists if the association between the nodesit connects is above the threshold and does not exist otherwise(21–23). However, the statistical validity of that method is ham-pered by the need to choose an arbitrary threshold as well as bythe discretization of inherently continuous edge weights. In thecurrent work, we have instead used fully weighted networks inwhich connections retain their original association value unlessthat value was found to be insignificant (based on statisticaltesting employing a false discovery rate correction for multiple

A B

C D

Fig. 4. Flexibility and learning. (A) Boxplots showing that the increase in flexibility from experimental session 1 to session 2 was significantly greater than zero(a one-sample t-test gives the result t " 6.00with p " 2 ! 10#8), and that the magnitude of the decrease in flexibility from session 2 to session 3 was significantlygreater than zero (t " 7.46, p " 2 ! 10#11). (B) Significant predictive correlations between flexibility in session 1 and learning in session 2 (black curve, p " 0.001)and between flexibility in session 2 and learning in session 3 (red curve, p " 0.009). Note that relationships between learning and network flexibility in the sameexperimental sessions (1 and 2) were not significant; we obtained p > 0.13 using permutation tests. (C) Brain regions whose flexibility in session 1 predictedlearning in session 2 (p < 0.05, uncorrected for multiple comparisons). Regions that also passed false-positive correction were the left anterior fusiform cortexand the right inferior frontal gyrus, thalamus, and nucleus accumbens. (D) Brain regions whose flexibility in session 2 predicted learning in session 3 (p < 0.05,uncorrected for multiple comparisons). Regions that also passed false-positive correction for multiple comparisons were the left intracalcarine cortex, para-cingulate gyrus, precuneus, and lingual gyrus and the right superior frontal gyrus and precuneus cortex. In (C) and (D), colors indicate the Spearman correlationcoefficient r between flexibility and learning.

7644 ! www.pnas.org/cgi/doi/10.1073/pnas.1018985108 Bassett et al.

Page 5: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

comparisons) (54). Future studies comparing multiple networkconstruction techniques will be important to statistically assessthe added value of weighted-edge retention in the assessmentof network correlates of cognition.

Second, partitioning a set of nodes into a set of communities isnondeterministic polynomial-time hard (55) so that modularity-optimization algorithms produce many near-optimal partitionsof the network (56). The number of near-optimal partitions tendsto be larger for large networks, and it also tends to be larger inbinary networks than in weighted ones (56). In the present paper,we study small weighted networks in which the number of near-optimal partitions is small. Nevertheless, we have systematicallyexplored the partition landscape in our optimization of the mod-ularity index. Accordingly, we report mean modularity estimatesthat our results suggest are representative (see SI Appendix).However, further work is necessary to measure common commu-nity assignments in the ensemble of partitions to identify consis-tently segregated groups of brain regions. Such research will aidin further exploration of the biological relevance of the detectedcommunities.

Finally, the statistical validation of community structure insocial and biological systems is complicated by several factors.For example, many investigations, especially in social systems,are hindered by their small number of instantiations. In our work,the relatively large number of subjects in conjunction with esti-mations of multiple networks over various temporal scales facili-tated a stringent statistical assessment of community structureboth in comparison to randomly connected graphs and, as wehave developed for dynamic networks, to graphs where nodalidentities or times were scrambled. An important future areaof research will focus on the development of alternative null mod-els that are not perfectly random but which assume increasinglybiologically realistic network architectures.

ConclusionConsistent with our hypotheses, we have identified significantmodular structure in human brain function during learningover a range of temporal scales: days, hours, and minutes. Mod-ular organization over short temporal scales changed smoothly,suggesting system adaptability. The composition of functionalmodules displayed temporal flexibility that was modulated byearly learning, varied over individuals, and was a significantpredictor of learning in subsequent experimental sessions.Furthermore, we developed and reported a general frameworkfor the statistical validation of dynamic modular architecturesin arbitrary systems. Additionally, our evidence for adaptive mod-ular organization in global brain activity during learning providescritical insight into the dependence of system performance onunderlying architecture.

Materials and MethodsTwenty-five right-handed participants (16 female, 9 male; mean age 24.25years) volunteered with informed consent in accordance with the Universityof California, Santa Barbara Internal Review Board. After exclusions for taskaccuracy, incomplete scans, and abnormal MRI, 18 participants were retainedfor subsequent analysis. All participants had less than 4 years of experiencewith any one musical instrument, had normal vision, and had no history ofneurological disease or psychiatric disorders. Participants were paid for theirparticipation.

The experimental framework consisted of a simple motor learning taskin which subjects responded to a visually cued sequence by generatingresponses using the four fingers of their nondominant hand on a custom re-sponse box. Participants were instructed to respond swiftly and accurately.Visual cues were presented as a series of musical notes on a pseudo-musicalstaff with four lines such that the top line of the staff mapped to the leftmostkey depressed with the pinkie finger. Each 12-note sequence contained threenotes per line, which were randomly ordered without repetition and free ofregularities such as trills and runs. The number and order of sequence trialswas identical for all participants. All participants completed three training

sessions in a five-day period, and each session was performed inside theMRI scanner.

Recordings with fMRI were conducted using a 3.0 T Siemens Trio with a12-channel phased-array head coil. For each functional run, a single-shotecho planar imaging sequence that is sensitive to blood oxygen level depen-dent (BOLD) contrast was used to acquire 33 slices (3 mm thickness) perrepetition time (TR), with a TR of 2,000 ms, an echo time of 30 ms, a flip angleof 90 °, a field of view of 192 mm, and a 64 ! 64 acquisition matrix. Imagepreprocessing was performed using the Oxford Center for FunctionalMagnetic Resonance Imaging of the Brain (FMRIB) Software Library (FSL),and motion correction was performed using FMRIB’s linear image registra-tion tool. Images were high-pass filtered with a 50 s cutoff period. Spatialsmoothing was performed using a kernel where full width at half maximumwas 8 mm. Signals were normalized globally to account for transient fluctua-tions in intensity.

The whole brain is parcellated into a set of N regions of interest thatcorrespond to the 112 cortical and subcortical structures anatomically iden-tified in FSL’s Harvard–Oxford atlas. For each individual fMRI dataset, we es-timate regional mean BOLD time series by averaging voxel time series in eachof the N regions. These regional time series are then subjected to a waveletdecomposition to reconstruct wavelet coefficients in the 0.06–0.12 Hz range(scale two). We estimate the correlation or coherence Aij between the activ-ity of all possible pairs of regions i and j to construct N ! N functionalconnectivity matrices A (Fig. 1A). Individual elements of Aij are subjectedto statistical testing, and the value of all elements that do not pass the falsediscovery rate correction for multiple comparisons are set to zero; other-wise, the values remain unchanged. The complete set of weighted networknodes is partitioned into communities by maximizing the modularity indexQ (30, 31). In the simplest static case, supposing that node i is assigned tocommunity gi and node j is assigned to community gj, the modularity indexis defined as

Q !!

ij

$Aij # Pij%!"gi;gj#; [1]

where !"gi;gj# ! 1 if gi ! gj and it equals 0 otherwise, and Pij is the expectedweight of the edge connecting node i and node j under a specified nullmodel. (A more complex formula is used in the dynamic network case;see SI Appendix.) The elements of the matrix Aij are weighted by the func-tional association between regions, and we thoroughly sample the distribu-tion of partitions that provide near-optimal Q values (56). The functionalconnectivity is termed “modular” if the value of Q is larger than thatexpected from random network null models that control for both the meanand variability of connectivity.

We tested for static modular structure on the individual networksand on dynamic network structure on a multilayer network created by link-ing networks between time steps (25). In both cases, we assess modularorganization using the modularity Q and the number of modules n. In thedynamic case, we also used two additional diagnostics to characterize mod-ular structure: the mean module size s and the stationarity of modules ".We defined s to be the mean number of nodes per community over all timewindows over which the community exists. We used the definition of modulestationarity from ref. 34. We started by calculating the autocorrelationfunction U"t;t &m# of two states of the same community G"t# atm time stepsapart using the formula

U"t;t&m#! jG"t# $ G"t&m#jjG"t#!G"t&m#j

; [2]

where jG"t# $ G"t &m#j is the number of nodes that are members of bothG"t# and G"t &m#, and jG"t#!G"t &m#j is the total number of nodes inG"t#!G"t &m# (34). We defined t0 to be the time at which a community isborn and t0 to be the final time step before a community is extinguished.The stationarity of a community is then

" !!

t0#1t!t0

U"t;t& 1#

t0 # t0 # 1; [3]

which is the mean autocorrelation over consecutive time steps (34).In principle, modular architecture might vary with learning by displaying

changes in global diagnostics such as the number of modules or the modu-larity index Q, or by displaying more specific changes in the compositionof modules. To measure changes in the composition of modules, we defined

Bassett et al. PNAS ! May 3, 2011 ! vol. 108 ! no. 18 ! 7645

SYST

EMSBIOLO

GY

APP

LIED

MAT

HEMAT

ICS

Page 6: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

the flexibility f i of a node to be the number of times that a node changedmodular assignment throughout the session, normalized by the total numberof changes that were possible (i.e., by the number of consecutive pairs oflayers in the multilayer framework). We then defined the flexibility F ofthe entire network as the mean flexibility over all nodes in the net-work: F ! 1

N%Ni!1 f i .

See SI Appendix for further mathematical details and methodologicaldescriptions.

ACKNOWLEDGMENTS. We thank two anonymous reviewers for helpful com-ments on this manuscript, Aaron Clauset for useful discussions, and JohnBushnell for technical support. This work was supported by the David andLucile Packard Foundation, Public Health Service Grant NS44393, the Institutefor Collaborative Biotechnologies through Contract W911NF-09-D-0001 fromthe US Army Research Office, and the National Science Foundation (Divisionof Mathematical Sciences-0645369). M.A.P. acknowledges research award220020177 from the James S. McDonnell Foundation.

1. Kim JJ, Thompson RF (1997) Cerebellar circuits and synaptic mechanisms involved inclassical eyeblink conditioning. Trends Neurosci 20:177–181.

2. Glanzman DL (2008) New tricks for an old slug: The critical role of postsynapticmechanisms in learning and memory in Aplysia. Prog Brain Res 169:277–292.

3. Xu T, et al. (2009) Rapid formation and selective stabilization of synapses for enduringmotor memories. Nature 462:915–919.

4. Büchel C, Coull JT, Friston KJ (1999) The predictive value of changes in effectiveconnectivity for human learning. Science 283:1538–1541.

5. Tunik E, Schmitt PJ, Grafton ST (2007) BOLD coherence reveals segregated functionalneural interactions when adapting to distinct torque perturbations. J Neurophysiol97:2107–2120.

6. Hart CB, Giszter SF (2010) A neural basis for motor primitives in the spinal cord.J Neurosci 30:1322–1336.

7. Simon HA (1962) The architecture of complexity. Proc Amer Philos Soc 106:467–482.8. Kirschner M, Gerhart J (1998) Evolvability. Proc Natl Acad Sci USA 95:8420–8427.9. Félix MA, Wagner A (1998) Robustness and evolution: Concepts, insights, and

challenges from a developmental model system. Heredity 100:132–140.10. Kashtan N, Alon U (2005) Spontaneous evolution of modularity and network motifs.

Proc Natl Acad Sci USA 102:13773–13778.11. Wagner GP, Altenberg L (1996) Complex adaptations and the evolution of evolvability.

Evolution 50:967–976.12. Schlosser G,Wagner GP, eds. (2004)Modularity in Development and Evolution (Univ of

Chicago, Chicago).13. Meunier D, Lambiotte R, Bullmore ET (2010) Modular and hierarchically modular

organization of brain networks. Front Neurosci 4:200, Available at http://www.frontiersin.org/neuroscience/10.3389/fnins.2010.00200/full.

14. Werner G (2010) Fractals in the nervous system: Conceptual implications for theore-tical neuroscience. Front Physiol 1:1–28.

15. Burdet E, Milner TE (1998) Quantization of human motions and learning of accuratemovements. Biol Cybern 78:307–318.

16. Sosnik R, Hauptmann B, Karni A, Flash T (2004) When practice leads to coarticulation:The evolution of geometrically defined movement primitives. Exp Brain Res156:422–438.

17. Schaal S, Schweighofer N (2005) Computational motor control in humans and robots.Curr Opin Neurobiol 15:675–682.

18. Doyon J, Benali H (2005) Reorganization and plasticity in the adult brain duringlearning of motor skills. Curr Opin Neurobiol 15:161–167.

19. Newell KM, Mayer-Kress G, Hong SL, Liu YT (2009) Adaptation and learning:Characteristic timescales of performance dynamics. Hum Mov Sci 28:655–687.

20. Lee JH, et al. (2010) Global and local fMRI signals driven by neurons defined optogen-etically by type and wiring. Nature 465:788–792.

21. Bullmore ET, Bassett DS (2010) Brain Graphs: Graphical models of the human brainconnectome. Annu Rev Clin Psychol 7:113–140.

22. Bassett DS, Bullmore ET (2009) Human brain networks in health and disease. Curr OpinNeurol 22:340–347.

23. Bassett DS, Bullmore ET (2006) Small-world brain networks.Neuroscientist 12:512–523.24. Bullmore ET, Sporns O (2009) Complex brain networks: Graph theoretical analysis of

structural and functional systems. Nat Rev Neurosci 10:186–198.25. Mucha PJ, Richardson T, Macon K, Porter MA, Onnela J-P (2010) Community structure

in time-dependent, multiscale, and multiplex networks. Science 328:876–878.26. Bassett DS, Meyer-Lindenberg A, Achard S, Duke T, Bullmore ET (2006) Adaptive

reconfiguration of fractal small-world human brain functional networks. Proc NatlAcad Sci USA 103:19518–19523.

27. Bassett DS, et al. (2010) Efficient physical embedding of topologically complexinformation processing networks in brains and computer circuits. PLoS Comput Biol6:e1000748.

28. Bassett DS, Brown JA, Deshpande V, Carlson JM, Grafton ST (2010) Conserved and vari-able architecture of human white matter connectivity. Neuroimage 54:1262–1279.

29. Masel J, Trotter MV (2010) Robustness and evolvability. Trends Genet 26:406–414.30. Porter MA, Onnela J-P, Mucha PJ (2009) Communities in networks. Not Am Math Soc

56:1082–1097 1164–1166.31. Fortunato S (2010) Community detection in graphs. Phys Rep 486:75–174.32. Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E (2008) Fast unfolding of community

hierarchies in large networks. J Stat Mech Theory Exp P10008.33. Maslov S, Sneppen K (2002) Specificity and stability in topology of protein networks.

Science 296:910–913.34. Palla G, Barabási A, Vicsek T (2007) Quantifying social group evolution. Nature

446:664–667.35. Meunier D, Achard S, Morcom A, Bullmore ET (2009) Age-related changes in modular

organization of human brain functional networks. Neuroimage 44:715–723.36. Meunier D, Lambiotte R, Fornito A, Ersche KD, Bullmore ET (2009) Hierarchical mod-

ularity in human brain functional networks. Front Neuroinformatics 3:37, Available athttp://www.frontiersin.org/neuroinformatics/10.3389/neuro.11.037.2009/full.

37. Damoiseaux J, Greicius MD (2009) Greater than the sum of its parts: A review ofstudies combining structural connectivity and resting-state functional connectivity.Brain Struct Funct 213:525–533.

38. Fodor JA (1983) Modularity of Mind: An Essay on Faculty Psychology (MIT Press,Cambridge, MA).

39. Raichle ME, Snyder AZ (2007) A default mode of brain function: A brief history of anevolving idea. Neuroimage 37:1083–1090.

40. Raichle ME (2010) Two views of brain function. Trends Cognit Sci 14:180–190.41. Achard S, Bassett DS, Meyer-Lindenberg A, Bullmore ET (2008) Fractal connectivity of

long-memory networks. Phys Rev E 77:036104.42. Hofer SB, Bonhoeffer T (2010) Dendritic spines: The stuff that memories are made of?

Curr Biol 20:R157–R159.43. Scholz J, Klein MC, Behrens TEJ, Johansen-Berg H (2009) Training induces changes in

white-matter architecture. Nat Neurosci 12:1370–1371.44. Smyth C, Summers JJ, Garry MI (2010) Differences in motor learning success are

associated with differences in M1 excitability. Hum Mov Sci 29:618–630.45. Ljubisavljevic M (2006) Transcranial magnetic stimulation and the motor learning-

associated cortical plasticity. Exp Brain Res 173:215–222.46. van Beers RJ (2009)Motor learning is optimally tuned to the properties of motor noise.

Neuron 63:406–417.47. Krakauer JW (2006) Motor learning: Its relevance to stroke recovery and neurorehabil-

itation. Curr Opin Neurol 19:84–90.48. Mulder T, Hochstenbach J (2001) Adaptability and flexibility of the human motor

system: Implications for neurological rehabilitation. Neural Plast 8:131–140.49. Faisal AA, Selen LP, Wolpert DM (2008) Noise in the nervous system. Nat Rev Neurosci

9:292–303.50. McIntosh AR, Kovacevic N, Itier RJ (2008) Increased brain signal variability accompanies

lower behavioral variability in development. PLoS Comput Biol 4:e1000106.51. Freeman WJ (1994) Characterization of state transitions in spatially distributed,

chaotic, nonlinear, dynamical systems in cerebral cortex. Integr Physiol Behav Sci29:294–306.

52. Deco G, Jirsa V, McIntosh AR, Sporns O, Kötter R (2009) Key role of coupling, delay, andnoise in resting brain fluctuations. Proc Natl Acad Sci USA 106:10302–10307.

53. Lippí S, Kovacevic N, McIntosh AR (2009) Differential maturation of brain signal com-plexity in the human auditory and visual system. Front HumNeurosci 3:48, Available athttp://www.frontiersin.org/human_neuroscience/10.3389/neuro.09.048.2009/full.

54. Genovese CR, Lazar NA, Nichols TE (2002) Thresholding of statistical maps in functionalneuroimaging using the false discovery rate. Neuroimage 15:870–878.

55. Brandes U, et al. (2008) On modularity clustering. IEEE T Knowl Data En 20:172–188.56. Good BH, de Montjoye YA, Clauset A (2010) Performance of modularity maximization

in practical contexts. Phys Rev E 81:046106.

7646 ! www.pnas.org/cgi/doi/10.1073/pnas.1018985108 Bassett et al.

Page 7: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Supplementary Material for

Dynamic reconfiguration of human brain networks during learning

Danielle S. Bassett1, Nicholas F. Wymbs2, Mason A. Porter3,4,

Peter J. Mucha5,6, Jean M. Carlson1, Scott T. Grafton2

1Complex Systems Group, Department of Physics, University of California, Santa Barbara, CA 93106,

USA

2Department of Psychology and UCSB Brain Imaging Center, University of California, Santa Barbara,

CA 93106, USA

3Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute, University of Oxford,

Oxford OX1 3LB, UK

4Complex Agent-Based Dynamic Networks Complexity Centre, University of Oxford, Oxford OX1 1HP,

UK

5Carolina Center for Interdisciplinary Applied Mathematics, Department of Mathematics, University of

North Carolina at Chapel Hill, NC 27599, USA

6Institute for Advanced Materials, Nanoscience & Technology, University of North Carolina, Chapel Hill,

NC 27599, USA

1

Page 8: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Contents

Full Description of Methods 3

Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Experimental Setup and Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Acquisition and Preprocessing of fMRI Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Partitioning the Brain into Regions of Interest . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Wavelet Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Connectivity over Multiple Temporal Scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Multilayer Network Modularity: Temporal Dynamics of Intra-Session Connectivity . . . . . . . 12

Temporal Dynamics of Brain Architecture and Learning . . . . . . . . . . . . . . . . . . . . . . 16

Statistics and Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Supplementary Results 18

Degeneracies of Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

Effect of Inter-Layer Coupling Parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

Effect of the Time Window Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

Learning and Flexibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

Supplementary Discussion 23

Resolution Limit of Modularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Measuring Differences in Brain States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

A Note on Computation Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2

Page 9: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Full Description of Methods

Sample

Twenty-five right-handed participants (16 female, 9 male) volunteered with informed consent in accor-

dance with the Institutional Review Board/Human Subjects Committee, University of California, Santa

Barbara. Handedness was determined by the Edinburgh Handedness Inventory. The mean age of the

participants was 24.25 years (range 18.5–30 years). Of these, 2 participants were removed because their

task accuracy was less than 60% correct, 1 was removed because of a cyst in presupplementary motor

area (preSMA), and 4 were removed for shortened scan sessions. This left 18 participants in total. All

participants had less than 4 years of experience with any one musical instrument, had normal vision, and

had no history of neurological disease or psychiatric disorders. Participants were paid for their partici-

pation. All participants completed 3 training sessions in a 5-day period, and each session was performed

inside the Magnetic Resonance Imaging (MRI) scanner.

Experimental Setup and Procedure

Participants were placed in a supine position in the MRI scanner. Padding was placed under the knees in

order to maximize comfort and provide an angled surface to position the stimulus response box. Padding

was placed under the left forearm to minimize muscle strain when participants typed sequences. Finally,

in order to minimize head motion, padded wedges were inserted between the participant and head coil of

the MRI scanner. For all sessions, participants performed a cued sequence production (CSP) task (see

Figure S1), responding to visually cued sequences by generating responses using their non-dominant (left)

hand on a custom fiber-optic response box. For some participants, a small board was placed between the

response box and the lap in order to help balance the box effectively. Responses were made using the 4

fingers of the left hand (the thumb was excluded). Visual cues were presented as a series of musical notes

on a 4-line music staff. The notes were reported in a manner that mapped the top line of the staff to the

leftmost key depressed with the pinkie finger and so on, so that notes found on the bottom line mapped

onto the rightmost key with the index finger (Figure S1B). Each 12-element note sequence contained 3

notes per line, which were randomly ordered without repetition and free of regularities such as trills (e.g.,

121) and runs (e.g., 123). The number and order of sequence trials was identical for all participants.

A trial began with the presentation of a fixation signal, which was displayed for 2 s. The complete

3

Page 10: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

12-element sequence was presented immediately following the removal of the fixation, and participants

were then instructed to respond as soon as possible. They were given a period of 8 s to type each

sequence correctly. Participants trained on a set of 16 unique sequences, and there were three different

levels of training exposure. Over the course of the three training sessions, three sequences—known as

skilled sequences—were presented frequently, with 189 trials for each sequence. A second set of three

sequences, termed familiar sequences, were presented for 30 trials each throughout training. A third set

composed of 10 different sequences, known as novice sequences, were also presented; each novice sequence

was presented 4–8 times during training.

Skilled and familiar sequences were practiced in blocks of 10 trials, so that 9 out of 10 trials were

composed of the same sequence and 1 of the trials contained a novice sequence. If a sequence was

reported correctly, then the notes were immediately removed from the screen and replaced with the

fixation signal, which remained on the screen until the trial duration (8 s) was reached. If there were any

incorrect movements, then the sequence was immediately replaced with the verbal cue INCORRECT and

participants subsequently waited for the start of the next trial. Trials were separated with an inter-trial

interval (ITI) lasting between 0 s and 20 s, not including any time remaining from the previous trial.

Following the completion of each block, feedback (lasting 12 s and serving as a rest) was presented that

detailed the number of correct trials and the mean time that was taken to complete a sequence. Training

epochs contained 40 trials (i.e., 4 blocks) and lasted a total of 345 scan repetition times (TRs), which

took a total of 690 s. There were 6 scan epochs per training session (2070 scan TRs). In total, each

skilled sequence was presented 189 times over the course of training (18 scan epochs; 6210 TRs).

In order to familiarize participants with the task, they were given a short series of warm-up trials

the day before the initial training session inside the scanner. Practice was also given in the scanner

during the acquisition of the structural scans and just prior to the start of the first training-session

epoch. Stimulus presentation was controlled with MATLAB R� version 7.1 (Mathworks, Natick, MA) in

conjunction with Cogent 2000 (Functional Imaging Laboratory, 2000). Key-press responses and response

times were collected using a fiber-optic custom button box transducer that was connected to a digital

response card (DAQCard-6024e; National Instruments, Austin, TX). We assessed learning using the slope

of the movement time (MT), which is the difference between the time of the first button press and the

time of the last button press in a single sequence (see Figure S1B) [1]. The negative slope of the movement

curve over trials indicates that learning is occurring [1].

4

Page 11: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Acquisition and Preprocessing of fMRI Data

Functional MRI (fMRI) recordings were conducted using a 3.0 T Siemens Trio with a 12-channel phased-

array head coil. For each functional run, a single-shot echo planar imaging that is sensitive to blood

oxygen level dependant (BOLD) contrast was used to acquire 33 slices (3 mm thickness) per repetition

time (TR), with a TR of 2000 ms, an echo time (TE) of 30 ms, a flip angle of 90 degrees, and a field of

view (FOV) of 192 mm. The spatial resolution of the data was defined by a 64 × 64 acquisition matrix.

Before the collection of the first functional epoch, a high-resolution T1-weighted sagittal sequence image

of the entire brain was acquired (TR = 15.0 ms, TE = 4.2 ms, flip angle = 9 degrees, 3D acquisition,

FOV = 256 mm; slice thickness = 0.89 mm, and spatial acquisition matrix dimensions = 256 × 256).

All image preprocessing was performed using the FMRIB (Oxford Centre for Functional Magnetic

Resonance Imaging of the Brain) Software Library (FSL) [2]. Motion correction was performed using

the program MCFLIRT (Motion Correction using FMRIB’s Linear Image Registration Tool). Images

were high-pass filtered with a 50 s cutoff period. Spatial smoothing was performed using a kernel where

the full width at half maximum was 8 mm. No temporal smoothing was performed. The signals were

normalized globally to account for transient fluctuations in signal intensity.

Partitioning the Brain into Regions of Interest

Brain function is characterized by a spatial specificity: different portions of the cortex emit inherently

different activity patterns that depend on the experimental task at hand. In order to measure the

functional connectivity between these different portions, it is common to apply an atlas of the entire

brain to raw fMRI data in order to combine information from all 3 mm cubic voxels found in a given

functionally or anatomically defined region (for recent reviews, see [3–5]). Several atlases are currently

available, and each provides slightly different parcellations of the cortex into discrete volumes of interest.

Several recent studies have highlighted the difficulty of comparing results from network analyses derived

from different atlases [6–8]. In the present work, we have therefore used a single atlas that provides the

largest number of uniquely identifiable regions—this is the Harvard-Oxford (HO) atlas, which is available

through the FSL toolbox [2,9]. The HO atlas provides 112 functionally and anatomically defined cortical

and subcortical regions; for a list of the brain regions, see Supplementary Table 1. Therefore, for each

individual fMRI data set, we estimated regional mean BOLD time series by averaging voxel time series

5

Page 12: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

in each of the 112 regions. Each regional mean time series was composed of 2070 time points for each of

the 3 experimental sessions (for a total of 6210 time points for the complete experiment).

Wavelet Decomposition

Brain function is also characterized by a frequency specificity; different cognitive and physiological func-

tions are associated with different frequency bands, which can be investigated using wavelets. Wavelet

decompositions of fMRI time series have been applied extensively in both resting-state and task-based con-

ditions [10,11]. In both cases, they provide increased sensitivity for the detection of small signal changes

in non-stationary time series with noisy backgrounds [12]. In particular, the maximum-overlap discrete

wavelet transform (MODWT) has been extensively used in connectivity investigations of fMRI [13–18].

Accordingly, we used MODWT to decompose each regional time series into wavelet scales correspond-

ing to specific frequency bands [19]. We were interested in quantifying high-frequency components of

the fMRI signal, correlations between which might be indicative of cooperative temporal dynamics of

brain activity during a task. Because our sampling frequency was 2 s (1 TR = 2 s), wavelet scale one

provided information on the frequency band 0.125–0.25 Hz and wavelet scale two provided information

on the frequency band 0.06–0.125 Hz. Previous work has indicated that functional associations between

low-frequency components of the fMRI signal (0–0.15 Hz) can be attributed to task-related functional

connectivity, whereas associations between high-frequency components (0.2–0.4 Hz) cannot [20]. This

frequency specificity of task-relevant functional connectivity is likely to be due at least in part to the

hemodynamic response function, which might act as a noninvertible bandpass filter on underlying neural

activity [20]. In the present study, we therefore restricted our attention to wavelet scale two in order to

assess dynamic changes in task-related functional brain architecture over short time scales while retaining

sensitivity to task-perturbed endogenous activity [21], which is most salient at about 0.1 Hz [22–24].

Connectivity Over Multiple Temporal Scales

Multiscale Connectivity Estimation We measured functional connectivity over three temporal

scales: the large scale of the complete experiment (which lasted 3 hours and 27 minutes), the session

time scale of each fMRI recording session (3 sessions of 69 minutes each; each session corresponded to

2070 time points), and the shorter time scales of intra-session time windows (where each time window

was approximately 3.5 min long and lasted 80 time points).

6

Page 13: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

In the investigation of large-scale connectivity, we concatenated regional mean time series over all 3

sessions, as has been done previously [25]. We then constructed for each subject a functional association

matrix based on correlations between regional mean time series. At the mesoscopic scale, we extracted

regional mean time series from each experimental session separately to compute session-specific matrices.

At the small scale, we constructed intra-session time windows with a length of T = 80 time points,

giving a total of 25 time windows in each session (see the Results section of this supplementary document

for a detailed investigation across a range of T values). We constructed separate functional association

matrices for each subject in each time window (25) for each session (3) for a total of 75 matrices per

subject. We chose the length of the time window to be long enough to allow adequate estimation of

correlations over the frequencies that are present in the wavelet band of interest (0.06–0.12 Hz), yet short

enough to allow a fine-grained measurement of temporal evolution over the full experiment.

Construction of Brain Networks To construct a functional network, we must first define a measure

of functional association between regions. Measures of functional association range from simple linear

correlation to nonlinear measures such as mutual information. In the majority of network investigations in

fMRI studies to date, the measure of choice has been the Pearson correlation [13,15,18,26,27], perhaps due

to its simplicity and ease of interpretation. Therefore, in order to estimate static functional association,

we calculated the Pearson correlation between the regional mean time series of all possible pairs of regions

i and j. This yields an N × N correlation matrix with elements ri,j , where N = 112 is the number of

brain regions of interest in the full brain atlas (see earlier section on “Partitioning the Brain into Regions

of Interest” for further details).

However, as pointed out in other network studies of fMRI data [13], not all elements ri,j of the full

correlation matrix necessarily indicate significant functional relationships. Therefore, in addition to the

correlation matrix element ri,j , we computed the p-value matrix element pi,j , which give the probabilities

of obtaining a correlation as large as the observed value ri,j by random chance when the true correlation

is zero. We estimated p-values using approximations based on the t-statistic using the MATLAB R�

function corrcoef [28]. In the spirit of Ref. [29] and following Ref. [13], we then tested the p-values pi,j

for significance using a False Discovery Rate (FDR) of p < 0.05 to correct for multiple comparisons [30,31].

We retained matrix elements ri,j whose p-values pi,j passed the statistical FDR threshold. Elements of

ri,j whose p-values pi,j did not pass the FDR threshold were set to zero in order to create new correlation

7

Page 14: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

matrix elements r�i,j .

We applied the statistical threshold to all ri,j independent of the sign of the correlation. Therefore,

the resulting r�i,j could contain both positive and negative elements if there existed both positive and

negative elements of ri,j whose p-values pi,j passed the FDR threshold. Because this was a statistical

threshold, the network density of r�i,j (defined as the fraction of non-zero matrix elements) was determined

statistically rather than being set a priori. Network density varied over temporal resolutions; the mean

density and standard deviation for networks derived from correlation matrices at the largest time scale

(3 hr and 27 minutes) was 0.906 (0.019%), at the intermediate time scale (69 min) was 0.846 (0.029), and

at the short time scale (3.5 min) was 0.423 (0.110).

We performed the procedure described above for each subject separately to create subject-specific

corrected correlation matrices. These statistically corrected matrices gave adjacency matrices A (see the

discussion below) whose elements were Aij = r�i,j .

Network Modularity To characterize the large-scale functional organization of the subject-specific

weighted matrices A, we used tools from network science [32]. In a network framework, brain regions

constitute the nodes of the network, and inter-regional functional connections that remain in the connec-

tivity matrix constitute the edges of the network. One powerful concept in the study of networks is that

of community structure, which can be studied using algorithmic methods [33, 34]. Community detection

is an attempt to decompose a system into subsystems (called ‘modules’ or ‘communities’). Intuitively, a

module consists of a group of nodes (in our case, brain regions) that are more connected to one another

than they are to nodes in other modules. A popular way to investigate community structure is to optimize

the partitioning of nodes into modules such that the quality function Q is maximized (see [33, 34] for

recent reviews and [35] for a discussion of caveats), for which we give a formula below.

From a mathematical perspective, the quality function Q is simple to define. One begins with a graph

composed of N nodes and some set of connections between those nodes. The adjacency matrix A is then

an N ×N matrix whose elements Aij detail a direct connection or ‘edge’ between nodes i and j, with a

weight indicating the strength of that connection. The quality of a hard partition of A into communities

(whereby each node is assigned to exactly one community) is then quantified using the quality function

Q. Suppose that node i is assigned to community gi and node j is assigned to community gj . The most

8

Page 15: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

popular form of the quality function takes the form [33,34]

Q =�

ij

[Aij − Pij ]δ(gi, gj) , (1)

where δ(gi, gj) = 1 if gi = gj and it equals 0 otherwise, and Pij is the expected weight of the edge

connecting node i and node j under a specified null model. (The specific choice of Q in Equation 1 is

called the network modularity or modularity index [36].) A most common null model (by far) used for

static network community detection is given by [33,34,37]

Pij =kikj

2m, (2)

where ki is the strength of node i, kj is the strength of node j, and m = 12

�ij Aij . The maximization

of the modularity index Q gives a partition of the network into modules such that the total edge weight

inside of modules is as large as possible (relative to the null model, subject to the limitations of the

employed computational heuristics, as optimizing Q is NP-hard [33,34,38]).

Network modularity has been used recently for investigations of resting-state functional brain networks

derived from fMRI [26,27] and of anatomical brain networks derived from morphometric analyzes [39]. In

these previous studies, brain networks were constructed as undirected binary graphs, so that each edge

had a weight of either 1 or 0. The characteristics of binary graphs derived from neuroimaging data are

sensitive to a wide variety of cognitive, neuropsychological, and neurophysiological factors [4,5]. However,

increased sensitivity is arguably more likely in the context of the weighted graphs that we consider, as

they preserve the information regarding the strength of functional associations (though, as discussed

previously, matrix elements ri,j that are statistically insignificant are still set to 0) [40]. An additional

contrast between previous studies and the present one is that (to our knowledge) investigation of network

modularity has not yet been applied to task-based fMRI experiments, in which modules might have a

direct relationship with goal-directed function.

We partitioned the networks represented by the weighted connectivity matrices into n communities by

using a Louvain greedy community detection method [41] to optimize the modularity index Q. Because

the edge weights in the correlation networks that we constructed contain both positive and negative

correlation coefficients, we used the signed null model proposed in Ref. [42] to account for communities of

9

Page 16: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

nodes associated with one another through both negative and positive edge weights. (Recall that we are

presently discussing aggregated correlation networks A, so we are detecting communities in single-layer

networks, as has been done in previous work. In order to investigate time-evolving communities, we will

later employ a new mathematical development that makes it possible to perform community detection in

multilayer networks [43].) We first defined w+ij to be an N ×N matrix containing the positive elements

of Aij and w−ij to be an N ×N matrix containing only the negative elements of Aij . The quality function

to be maximized is then given by

Q± =1

2w+ + 2w−

i

j

�Aij −

�γ+

w+i w+

j

2w+− γ−

w−i w−

j

2w−

��δ(gigj) , (3)

where gi is the community to which node i is assigned, gj is the community to which node j is assigned,

γ+ and γ− are resolution parameters, and w+i =

�j w+

ij , w−i =

�j w−

ij [42]. For simplicity, we set the

resolution parameter values to unity.

In our investigation, we have focused on the mean properties of ensembles of partitions rather than

on detailed properties of individual partitions. This approach is consistent with recent work illustrating

the fact that the optimization of quality functions like Q and Q± is hampered by the complicated shape

of the optimization landscape. In particular, one expects to find a large number of partitions with near-

optimum values of the quality function [35], collectively forming a high-modularity plateau. Theoretical

work estimates that the number of “good” (in the sense of high values of Q and similar quality functions)

partitions scales as 2n−1, where n is the mean number of modules in a given partition [35]. In both toy

networks and networks constructed from empirical data, many of the partitions found by maximizing

a quality function disagree with one another on the components of even the largest module, impeding

interpretations of particular partitions of a network [35]. Therefore, in the present work, we have focused

on quantifying mean qualities of the partitions after extensive sampling of the high-modularity plateau.

Importantly, the issue of extreme near-degeneracy of quality functions like Q is expected to be much less

severe in the networks that we consider than is usually the case, because we are examining small, weighted

networks rather than large, unweighted networks [35]. We further investigate the degenerate solutions in

terms of their mean, standard deviation, and maximum. We find that Q± values are tightly distributed,

with maximum values usually less than three standard deviations from the mean (see Supplementary

Results).

10

Page 17: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Statistical Testing To determine whether the value of Q± or the number of modules was greater

or less than expected in a random system, we constructed randomized networks with the same degree

distribution as the true brain networks. As has been done previously [27,44], we began with a real brain

network and then iteratively rewired using the algorithm of Maslov and Sneppen [45]. The procedure we

used for accomplishing this rewiring was to choose at random two edges—one that connects node A to

node B and another that connects nodes C and D—and then to rewire them to connect A to C and B

to D. This allows us to preserve the degree, or number of edges, emanating from each node although it

does not retain a node’s strength [in such weighted networks]. To ensure a thorough randomization of the

underlying connectivity structure, we performed this procedure multiple times, such that the expected

number of times that each edge was ‘rewired’ was 20. This null model will be hereafter referred to as

the static random network null model. (This is distinct from the null models that we have developed for

statistical testing of community structure in multilayer networks, as discussed in the main manuscript

and in later sections of this Supplement.) The motivation for this process is to compare the brain with a

null model that resembles the configuration model [46], which is a random graph with prescribed degree

distribution.

We constructed 100 instantiations of the static random network null model for each real network that

we studied. We constructed representative values for diagnostics from the random networks by taking

the mean network modularity and mean number of modules over those 100 random networks. We then

computed the difference between the representative random values and the real values for each diagnostic,

and we performed a one-sample t-test over subjects to determine whether that difference was significantly

greater than or less than zero. For each case, we then reported p-values for these tests.

Sampling of the static random network null model distribution is important in light of the known

degeneracies of modularity (which we discuss further in the Supplementary Results section below) [35].

One factor that accounts for a significant amount of variation in Q± is the size (i.e., number of nodes)

of the network, so comparisons between networks of different sizes must be performed with caution.

Therefore, we note that all networks derived from the aforementioned null model retain both the same

number of nodes and the same number of edges as the real networks under study. This constrains

important factors in the estimation of Q±.

While the results reported in the main manuscript (see Figure 2) are based on the above mentioned

random network null model which preserves the degree distribution of the empirical networks (hereafter

11

Page 18: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

RD), we also report here results for a random network null model which separately preserves the degree

distribution of the matrix of positive elements w+ij and that of the matrix of negative elements w−

ij

(hereafter RWW). To construct an ensemble of RWW networks, we use a slightly altered version of

the rewiring algorithm of Maslov and Sneppen described above. First, with equal probability we choose

either two positive edges or two negative edges (which connect nodes A to B and nodes C to D). We

then rewire these edges to link nodes A to C and nodes B to D if and only if no link (whether positive or

negative) exists between either A and C or B and D. We find that for the large temporal scale of the entire

experiment, the modularity of RWW networks was significantly lower than that of RD networks (two-

sample t-test over subjects: t ≈ 2.29, p ≈ 0.02), but the number of modules was unchanged (t ≈ 1.81,

p ≈ 0.07). Similarly for the intermediate temporal scale of the 3 experimental sessions, we found that the

modularity of RWW networks was significantly lower than that of RD networks (t ≈ 3.31, p ≈ 0.002),

but the number of modules was unchanged (t ≈ 1.20, p ≈ 0.23). For the smallest spatial scale of

individual time windows, we found no difference between the results generated by the two null models:

for modularity, t ≈ 0.64 and p ≈ 0.52, while for the number of modules, t ≈ 0.27 and p ≈ 0.77. These

results indicate that for the two larger temporal scales, the results of the RD null model reported in

the main manuscript allow us to make a conservative estimate of the modularity differences between the

cortical structure and a random structure. For the smallest temporal scale, the choice of null model does

not statistically alter our results.

Visualization of Networks We visualized networks using the software package MATLAB R� (2007a,

The MathWorks Inc., Natick, MA). Following Ref. [47], we used the Fruchtermann-Reingold algorithm [48]

to determine node placement for a given network with respect to the extracted communities and then

used the Kamada-Kawai algorithm [49] to place the nodes within each community.

Multilayer Network Modularity: Temporal Dynamics of Intra-Session Con-

nectivity

In order to investigate the temporal evolution of modular architecture in human functional connectivity,

we used a mutilayer network framework in which each layer consists of a network derived from a single

time window. Networks in consecutive layers therefore correspond to consecutive time windows. We

linked networks in consecutive time windows by connecting each node in one window to itself in the

12

Page 19: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

previous and in the next windows (as shown in Figure 3A-B in the main text) [43]. We constructed a

multilayer network for each individual and in each of the three experimental sessions. We then performed

community detection by optimizing a multilayer modularity (see the discussion below) [43] using the

Louvain greedy algorithm (suitably adapted for this more general structure) on each multilayer network

in order to assess the modular architecture in the temporal domain.

In our examination of static network architecture, we used the wavelet correlation to assess func-

tional connectivity. Unfortunately, more sensitive measures of temporal association such as the spectral

coherence are not appropriate over the long time scales assessed in the static investigation due to the

nonstationarity of the fMRI time series [10–12], and it is exactly for this reason that we have used the

wavelet correlation for the investigation of aggregated (static) networks. However, over short temporal

scales such as those being used to construct the multilayer networks, fMRI signals in the context of the

motor learning task that we study can be assumed to be stationary [50], so spectral measures such as the

coherence are potential candidates for the measurement of functional association.

In the examination of the dynamic network architecture of brain function using multilayer community

detection, our goal was to measure temporal adaptivity of modular function over short temporal scales.

In order to estimate that temporal adaptivity with enhanced precision, we used the magnitude squared

spectral coherence (as estimated using the minimum-variance distortionless response method [51]) as a

measure of nonlinear functional association between any two time series. In using the coherence, which

has been demonstrated to be useful in the context of fMRI neuroimaging data [20], we were able to

measure frequency-specific linear relationships between time series.

As in the static network analysis described earlier, we tested the elements of each N ×N coherence

matrix (which constitutes a single layer) for significance using an FDR correction for multiple comparisons.

We used the original weighted (coherence values) of network links corresponding to the elements that

passed this statistical test, while those corresponding to elements that did not pass the test were set to

zero. In applying a community detection technique to the resulting coherence matrices, it is important to

note that the coherence is bounded between 0 and 1. We can therefore use a multilayer quality function

with an unsigned null model rather than the signed null model used in the static case described earlier.

The multilayer modularity Qml is given by [43]

Qml =12µ

ijlr

��Aijl − γl

kilkjl

2ml

�δlr + δijCjlr

�δ(gil, gjr) , (4)

13

Page 20: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

where the adjacency matrix of layer l (i.e., time window number l) has components Aijl, γl is the resolution

parameter of layer l, gil gives the community assignment of node i in layer l, gjr gives the community

assignment of node j in layer r, Cjlr is the connection strength between node j in layer r and node j in

layer l (see the discussion below), kil is the strength of node i in layer l, 2µ =�

jr κjr, κjl = kjl +cjl, and

cjl =�

r Cjlr. For simplicity, as in the static network case, we set the resolution parameter γl to unity

and we have set all non-zero Cjlr to a constant C, which we will term the ‘inter-layer coupling’. In the

main manuscript, we report results for C = 1. In the Supplementary Results section of this document,

we investigate the dependence of our results on alternative choices for the value of C.

Diagnostics We used several diagnostics to characterize dynamic modular structure. These include

the multilayer network modularity Qml, the number of modules n, the module size s, and the stationarity

of modules ζ. We defined the size of a module s to be the mean number of nodes per module over all time

windows over which the community exists. We used the definition of module stationarity from Ref. [52].

We started by calculating the autocorrelation function U(t, t + m) of two states of the same community

G(t) at m time steps apart using the formula

U(t, t + m) ≡ |G(t) ∩G(t + m)||G(t) ∪G(t + m)| , (5)

where |G(t) ∩ G(t + m)| is the number of nodes that are members of both G(t) and G(t + m), and

|G(t) ∪G(t + m)| is the total number of nodes in G(t) ∪G(t + m) [52]. We defined t0 to be the time at

which the community is born and t� to be the final time step before the community is extinguished. The

stationarity of a community is then

ζ ≡�t�−1

t=t0U(t, t + 1)

t� − t0 − 1, (6)

which is the mean autocorrelation over consecutive time steps [52].

Statistical Framework The study of the “modular architecture” of a system is of little value if the

system is not modular. It is therefore imperative to statistically quantify the presence or absence of

modular architecture to justify the use of community detection in a given application. Appropriate

random null models have been developed and applied to the static network framework [27, 44], but no

such null models yet exist for the multilayer framework. We therefore developed several null models in

14

Page 21: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

order to statistically test the temporal evolution of modular structure. We constructed three independent

null models to test for (1) network structure dependent on the topological architecture of connectivity,

(2) network structure dependent on nodal identity, and (3) network structure dependent on the temporal

organization of layers in the multilayer framework.

In the connectional null model (1), we scrambled links between nodes in any given time window (the

entire experiment, 3.45 hr; the individual scanning session, 69 min; or intra-session time windows, 3.45

min) while maintaining the total number of connections emanating from each node in the system. To

be more precise, for each layer of the multilayer network, we sampled the static random network null

model (see the discussion above in the context of static connectivity architecture) for that particular

layer. That is, we reshuffled the connections within each layer separately while maintaining the original

degree distribution. We then linked these connectivity-randomized layers together by coupling a node in

one layer to itself in contiguous layers to create the connectional null model multilayer network, just as we

connected the real layers to create the real multilayer network. In the present time-dependent context,

we performed this procedure on each time window in the multilayer network, after which we applied

the multilayer community detection algorithm to determine the network modularity of the randomized

system.

In constructing a nodal null model (2), we focused on the links that connected a single node in one

layer of the multilayer framework to itself in the next and previous layers. In the null model, the links

between layers connect a node in one layer to randomly-chosen nodes in contiguous layers instead of

connecting the node to itself in those layers. Specifically, in each time window τi (except for the final

one), we randomly connected the nodes in the corresponding layer to other nodes in the next time window

(τi+1) such that no node in τi was connected to more than one node in τi+1. We then connected nodes

in τi+1 to randomly-chosen nodes in τi+2, and so on until links between all time windows had been fully

randomized.

We also considered randomization of the order in which time windows were placed in the multilayer

network to construct a temporal null model (3). In the real multilayer construction, we (of course)

always placed the network from time window τi just before the network from time window τi+1. In

the temporal null model, we randomly permuted the temporal location of the individual layer in the

multilayer framework such that the probability of any time window τi following any other time window

τ, (j �= i) was uniform.

15

Page 22: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Statistical Testing In both the real network and the networks derived from null models (1)–(3), it is

important to adequately sample the distributions of partitions meant to optimize the modularity index

Qml. This step in our investigation was particularly important in light of the extreme degeneracy of the

network modularity function Qml [35] (see the Supplementary Results section on the Degeneracies of Q

for a quantitative characterization of such degeneracies).

Because the multilayer community detection algorithm can find different maxima each time it is run,

we computed the community structure of each individual real multilayer network a total of 100 times.

We then averaged the values of all diagnostics (modularity index Qml, number of modules, module size,

and stationarity) over those 100 partitions to create a representative real value. To perform our sampling

for the null models, we considered 100 multilayer network instantiations for each of the three different

null models. We also performed community detection on these null models using our multilayer network

adaptation of the Louvain modularity-optimization algorithm [41] to create a distribution of values for

each diagnostic. We then used the mean value of each diagnostic in our subsequent investigation as the

representative value of the null model.

We used one-sample t-tests to test statistically whether the differences between representative values

from the real networks and null model networks over the subject population was significantly different

from zero. The results, which we reported in the main manuscript, indicated that in contrast to what we

observed using each of the three null models, the human brain displayed a heightened modular structure.

That is, it is composed of more modules, which have smaller sizes. Considering the three null models in

order, this suggests that cortical connectivity has a precise topological organization, that cortical regions

consistently maintain individual connectivity signatures necessary for cohesive community organization,

and that functional communities evolve cohesively in time (see Figure 2 in the main manuscript). Im-

portantly, the stationarity of modular organization ζ was also higher in the human brain than in the

connectional or nodal null models, indicating a cohesive temporal evolution of functional communities.

Temporal Dynamics of Brain Architecture and Learning

In the present study, we have attempted to determine whether changes in the dynamic modular architec-

ture of functional connectivity is shaped by learning. We assessed the learning in each session using the

slope of the movement times (MT) of that session. Movement time is defined as the difference between

the time of the first button press and the time of the last button press in a single sequence (see Figure

16

Page 23: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

S1B). During successful learning, movement time is known to fall logarithmically with time [1]. However,

two subjects from session 1 and one subject from session 2 showed an increasing movement time as the

session progressed. We therefore excluded these three data points in subsequent comparisons due to the

decreased likelihood that successful learning was taking place. This process of screening participants

based on movement time slope is consistent with previous work suggesting that fMRI activation patterns

during successful performance might be inherently different when performance is unsuccessful [53].

In principle, modular architecture might vary with learning by displaying changes in global diagnostics

such as the number of modules or the modularity index Q or by displaying more specific changes in the

composition of modules. To measure changes in the composition of modules, we defined the flexibility

of a node fi to be the number of times that node changed modular assignment throughout the session,

normalized by the total number of changes that were possible (i.e., by the number of consecutive pairs

of layers in the multilayer framework). We then defined the flexibility of the entire network as the mean

flexibility over all nodes in the network: F = 1N

�Ni=1 fi.

Statistics and Software

We implemented all computational and simple statistical operations using the software packages MATLAB R�

(2007a, The MathWorks Inc., Natick, MA) and Statistica R� (version 9, StatSoft Inc.). We performed the

network calculations using a combination of in-house software (including multilayer community detection

code [43]) and the Brain Connectivity Toolbox [40].

17

Page 24: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Supplementary Results

Degeneracies of Q

As discussed earlier in the Methods section, we focused in this investigation on the mean properties of

ensembles of partitions rather than on detailed properties of individual partitions. Our approach was

motivated by recent work indicating that the optimization of modularity and similar quality functions

is hampered by the complicated shape of the optimization landscape, which includes a large number of

partitions with near-optimum values that collectively form a high modularity plateau [35]. To quantify

and address this degeneracy of Q± and Qml, we now provide supplementary results on the mean, standard

deviation, and maximum values of Q± and Qml over the 100 samples of the plateau computed for all real

networks in both the static and dynamic frameworks.

The mean number of modules in a given partition in the static framework was n ≈ 3.08 for the entire

experiment, n ≈ 3.07 for individual experimental sessions, and n ≈ 3.55 for the small intra-session time

windows. The mean number of modules in a given partition in the multilayer framework was n ≈ 6.00.

We have therefore chosen to sample the quality functions Q± and Qml a total of 100 times (which is more

than 2n−1 in each case, and therefore adequately samples the degenerate near-optimum values of Q± and

Qml [35]). In order to characterize the distribution of solutions found in these 100 samplings, we have

computed the mean, standard deviation, and maximum of Q± (static cases) and Qml (dynamic cases);

see Figure S2. We found that the values of Q± and Qml are tightly distributed, and that the maximum

values of Q± or Qml are between 0 and 3 standard deviations higher than the mean. Although we

remain cautious because we have not explored all possible computational heuristics, we are nevertheless

encouraged by these results that the mean values of Q± and Qml that we have reported are representative

of the true maximization of the two quality functions.

Reproducibility We calculated the intra-class correlation coefficient (ICC), to determine whether

values of Q± and Qml derived from a single individual over the 100 samples were more similar than

values of Q± or Qml derived from different individuals. The ICC is a measure of the total variance for

which between-subject variation accounts [54,55], and it is defined as

ICC =σ2

bs

σ2bs + σ2

ws

, (7)

18

Page 25: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

where σbs is the between-subject variance and σws is the pooled within-subject variance (‘pooled’ indicates

that variance was estimated for each subject and then averaged over subjects). The ICC is normalized

to have a maximum value of 1; values above 0.5 indicate that there is more variability between Q± and

Qml values from different subjects than between Q± and Qml values from the same subject. In the static

framework, the ICC was 0.9884 at the large scale (the entire experiment), an average of 0.9863 at the

intermediate scale (three experimental sessions), and an average of 0.9847 at the small scale (individual

time windows). In the multilayer framework, we calculated that ICC ≈ 0.9983. These results collectively

indicate that the Q± and Qml values that we reported in this work were significantly reproducible over

the 100 samples of the respective quality function landscape. That is, the Q± or Qml values drawn from

the 100 samples of a single subject’s network modularity landscape were more similar than Q± or Qml

values drawn from different subjects.

Effect of the Inter-Layer Coupling Parameter

The multilayer network framework requires one to define a coupling parameter C that indicates the

strength of the connections from a node in one time window to itself in the two neighboring time windows

[43]. In order to be sensitive to both temporal dynamics and intra-layer network architecture, the coupling

parameter should be on the same scale of values as the edge weights. For example, if edge weights are

coherence values lying between 0 and 1, then the coupling parameter also ought to lie between 0 and 1.

In the results that we presented in the main manuscript, we set the coupling parameter to be C = 1,

which is the highest value consistent with the intra-layer edge weights given by the normalized coherence.

However, if we were to alter the coupling value, one might expect the number of communities to be

altered in kind. As the strength of the coupling is increased, one might expect fewer communities to be

uncovered due to the increased temporal dependence between layers [43]. Similarly, as the inter-layer

coupling is weakened, one might expect more communities to be detected.

To probe the effect of the inter-layer coupling strength, we thus varied C from well below to well above

the maximum intra-layer edge weight (0.2 ≤ C ≤ 2). In Figure S3 (cortical network results are shown

in blue), we illustrate the effects of sweeping over this coupling parameter on our four diagnostics. The

modularity index Qml increases with increasing inter-layer coupling, whereas the other three diagnostics—

number of modules, module size, and stationarity—increase initially and then plateau approximately at

about C = 1 and above. The change in behavior near C = 1 can be rationalized as follows: For C < 1,

19

Page 26: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

intra-layer edge weights dominate the modularity optimization, whereas inter-layer edge weights dominate

for C > 1. The proposed choice of C = 1 therefore balances the impact of known coherence in brain

activity (as given by the intra-layer edge weights) on measured architectural adaptations and is therefore

a natural choice with which to investigate biologically meaningful organization.

We also computed 100 temporal, nodal, and connectional null model networks for each of the additional

coupling parameter values (see Figure S3; null model network results shown in green, orange, and red).

The results indicate that the relationship between diagnostics in the cortical networks and null model

networks is dependent on the diagnostic. For example, modularity values of null model networks are

consistently lower than modularity values of cortical networks. However, stationarity in the null model

networks is lower than that in cortical networks for small values of C but higher than that of cortical

networks for high values of C. This nontrivial behavior suggests an added sensitivity of the proposed

null model networks to the multilayer network construction, which might be useful in other experimental

contexts and therefore warrants further investigation.

Effect of the Time Window Length

In the construction of networks at the smallest time scale, it is necessary to choose a length of the time

window T . In choosing this time window length, two considerations are important: (1) the time window

must be short enough to adequately measure temporal evolution of network structure, and (2) the time

window must be long enough to adequately estimate the functional association between two time series

using (for example) the correlation or coherence [56]. In the main text, we reported results for time

windows of 80 data points in length. This gives 25 time windows in each experimental session, for a

total of 75 time windows over the 3 sessions. In addition to this extensive coverage of the underlying

temporal dynamics, the choice of a time window of 80 data points in length also ensures that 20 data

points can be used for the estimation of the functional association between time series in the frequency

band of interest—i.e., at wavelet scale two (0.06–0.12 Hz). If one were to increase the time window

length, one would expect a decreased ability to measure temporal variations due to the presence of fewer

time windows per session. If one were to decrease the time window length, one would expect increased

variance in the estimation of the functional association between time series due to the use of fewer data

points in the estimation of either the coherence or the correlation [16].

To probe the effect of the time window length, we varied T from T = 80 to T = 110 (see Figure

20

Page 27: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

S4; cortical network results are shown in blue). We find that the stationarity of the modules increases

with increasing time window length. As T is increased, the functional association between any two nodes

is averaged over a longer time series, so small adaptations over shorter time scales can no longer be

measured. This smoothing is likely the cause of the increased stationarity that we find at high values of

T . It suggests that functional association measured over long time windows is less dependent on the time

window being used than functional association measured over short time windows. This finding supports

our choice of short time windows in order to measure dynamic adaptations in network architecture.

We also computed 100 temporal, nodal, and connectional null model networks for each of the additional

time window lengths (see Figure S4; null model network results are shown in green, orange, and red).

The results indicate that the relationships between diagnostics in the cortical networks and null model

networks are largely conserved across time window lengths.

Learning and Flexibility

In the main text, we reported a significant correlation between the flexibility of dynamic modular archi-

tecture in a given experimental session, as measured by the (normalized) number of times a node changes

module allegiance, and learning in the subsequent experimental sessions, as measured by the slope of

the movement time (see Methods). We found that the mean value of flexibility was approximately 0.30,

that it fluctuated over the three experimental sessions, and that the values were highest in the second

experimental session (see Table 2 in this Supplement). We followed this large-scale calculation with an

investigation into the relationship between nodal flexibility (in particular brain regions) and learning. We

found, as shown Figure 4 of the main manuscript, that the flexibility of a large number of brain regions

could be used to predict learning in the following session. Here we also note that these regions were not

those with highest flexibility or lowest flexibility in the brain. In fact, the flexibility of those regions that

predicted learning was not significantly different from the flexibility of those regions that did not predict

learning: t ≈ 0.01 p ≈ 0.98 (Session 1) and t ≈ 0.87, p ≈ 0.38 (Session 2).

In addition to those results reported in the main manuscript, we tested whether the flexibility of

the cortical networks was significantly different from the flexibility expected in the (connectional, nodel,

and temporal) random network null models. As we show in Table 3 in this Supplement, the flexibility

of the connectional and nodal null model networks was significantly higher than that of the cortical

networks, and we found no discernible differences between the cortical networks and the temporal null

21

Page 28: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

model networks. We found the greatest degree of flexibility in the nodal null model, in which individual

nodes in any given time window were coupled to randomly selected nodes in the following time window.

It is thus plausible that the subsequent disruption of nodal identity caused nodes to change computed

module allegiances in this null model.

Robustness to Alternative Definitions It is important to assess the robustness of our findings

to different definitions of flexibility. We therefore defined an alternative flexibility measure f �i to be

the number of communities (modules) to which a node belongs at some point in a given experimental

session. The mean alternative flexibility F � is then given by averaging f �i over all nodes in the network:

F � = 1N

�Ni=1 f �i . Using this alternative definition of flexibility, we again tested for differences between the

cortical network and the three random network null models. As shown in Tables 2 and 3, the F � values

of cortical networks were also significantly different from those in the null model networks. Interestingly,

for this alternative definition, the temporal network null model exhibits significantly lower flexibility

than the cortical networks, suggesting that this measure of flexibility might be sensitive to biologically

relevant temporal evolution of modular architecture. Finally, we tested whether this alternative definition

of flexibility also displayed a relationship to learning. Flexibility and learning were not significantly

correlated in Session 1 (r ≈ 0.02, p ≈ 0.90) or in Session 2 (r ≈ 0.18, p ≈ 0.48), but flexibility in Session 1

was predictive of learning in Session 2 (r ≈ 0.64, p ≈ 0.002), and flexibility in Session 2 was predictive of

learning in Session 3 (r ≈ 0.51, p ≈ 0.019). These results for the alternative flexibility F � are consistent

with those of the original definition F , suggesting that our findings are robust.

22

Page 29: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Supplementary Discussion

Resolution Limit of Modularity

When detecting communities by optimizing modularity and similar quality functions, it is important to

note that modularity suffers from a resolution limit [33–35, 57]. As a result, the maximum-modularity

partition can be biased towards a particular module size and can have difficulty resolving modules smaller

than that size. Consequently, small modules of potential interest have the potential to be hidden within

larger groups of nodes that have been detected. Modularity’s resolution limit is particularly prevalent in

sparse networks, binary networks, and large networks, and its effects tend to be much less significant in

networks of the type (dense, weighted, and small) that we have studied [35].

Measuring Differences in Brain States

In the present work, we have characterized differences in brain states during learning by examining the

global network architecture and measuring differences between that architecture over three experimental

sessions. An alternative line of investigation would be to seek network motifs (i.e., small patterns of nodes

and edges) that have the potential to distinguish between brain states. This could be done using statistical

methods [58], machine-learning techniques [59], or a combination of the two [60]. Our approach, however,

has the advantage of assessing alterations in large-scale achitectural properties rather than differences

in small parts of that architecture. Additionally, the approach that we have chosen provides a direct

characterization of the underlying functional connectivity architecture irrespective of differences between

brain states. Using this approach, we have therefore been able to demonstrate, for example, that there

is significant non-random modular organization across multiple temporal scales.

A Note on Computational Time

The investigations that we reported in the present work involved about 10, 000 CPU-days, and our

study was therefore made possible by the use of two computing clusters available at the Institute for

Collaborative Biotechnologies at UC Santa Barbara. Cluster 1 was composed of 42 Dell SC1425s (dual

single-core Xeon 2.8GHz, 4GB memory), 5 Dell PE1950s (dual quad-core Xeon E5335 2.0GHz, 8GB

memory), 1 Dell 2850 (RAID storage includes 500GB for the home directory), and MATLAB R� MDCE

with 128 worker licenses (cluster currently has 124 compute cores), Gigabit Ethernet, Software RAID

23

Page 30: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

backup node (converted compute node) with 673GB software RAID backup. Cluster 2 was composed of

20 HP Proliant DL160 G6s (dual quad-core E5540 “Nehalem” 2.53GHz, 24GB memory), 1 HP DL180

G6 (RAID storage includes 2.1TB for the home directory), MATLAB R� MDCE with 160 worker licenses

(cluster currently has 160 compute cores), Gigabit Ethernet, and a storage node with 4.6TB of RAID

storage (for backup).

We performed maximization of the quality functions (Qpm, Qml) a total of 100 times for every

connectivity matrix under study. In the static connectivity investigation, we constructed connectivity

matrices for 20 subjects, 3 temporal scales (encompassing 1 experiment, 3 experimental sessions, and

25 time windows), and 1 random network null model. In the dynamic connectivity investigation, we

constructed connectivity matrices for 20 subjects, 18–34 time windows, 3 different null models, 10 values

of the inter-layer coupling C, and 4 values of time window length (80, 90, 100, and 110 TRs). In light

of the computational extent of this work, we note that we did not employ Kernighan-Lin (KL) node-

swapping steps [61] in our optimization of Qpm or Qml, as they would be computationally prohibitive and

are not necessary in the present context. KL steps move individual nodes between communities in order

to further optimize a single sample of Qpm or Qml [33, 62, 63]. As we focus on the mean properties of

ensembles of partitions (and use them to report reliable measurements of architectural properties) rather

than on the values of diagnostics for any individual partitions, KL steps that provide a marginal increase

in the value of Qpm or Qml would not be helpful for our study.

24

Page 31: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Frontal pole Cingulate gyrus, anteriorInsular cortex Cingulate gyrus, posteriorSuperior frontal gyrus Precuneus cortexMiddle frontal gyrus Cuneus cortexInferior frontal gyrus, pars triangularis Orbital frontal cortexInferior frontal gyrus, pars opercularis Parahippocampal gyrus, anteriorPrecentral gyrus Parahippocampal gyrus, posteriorTemporal pole Lingual gyrusSuperior temporal gyrus, anterior Temporal fusiform cortex, anteriorSuperior temporal gyrus, posterior Temporal fusiform cortex, posteriorMiddle temporal gyrus, anterior Temporal occipital fusiform cortexMiddle temporal gyrus, posterior Occipital fusiform gyrusMiddle temporal gyrus, temporooccipital Frontal operculum cortexInferior temporal gyrus, anterior Central opercular cortexInferior temporal gyrus, posterior Parietal operculum cortexInferior temporal gyrus, temporooccipital Planum polarePostcentral gyrus Heschl’s gyrusSuperior parietal lobule Planum temporaleSupramarginal gyrus, anterior Supercalcarine cortexSupramarginal gyrus, posterior Occipital poleAngular gyrus CaudateLateral occipital cortex, superior PutamenLateral occipital cortex, inferior Globus pallidusIntracalcarine cortex ThalamusFrontal medial cortex Nucleus AccumbensSupplemental motor area Parahippocampal gyrus (superior to ROIs 34,35)Subcallosal cortex HippocampusParacingulate gyrus Brainstem

Table 1: Brain regions present in the Harvard-Oxford Cortical and Subcortical Parcellation Scheme pro-vided by FSL [2,9].

Type of Flexibility Session Cortical Connectional Nodal TemporalF

1 0.027±0.009 0.041±0.008 0.070±0.001 0.027±0.0102 0.030±0.009 0.042±0.007 0.067±0.001 0.030±0.0093 0.027±0.008 0.039±0.008 0.064±0.001 0.026±0.009

F �

1 1.93±0.30 2.58±0.25 3.19±0.01 1.90±0.312 1.95±0.27 2.53±0.23 3.01±0.01 1.94±0.283 1.86±0.25 2.43±0.23 2.92±0.01 1.85±0.26

Table 2: Flexibility in cortical network and random network null models for two different definitions of flexibility.

One definition is F =1N

PNi=1 fi, where fi is the number of times node i changes module allegiance in a given experimental

session divided by the total number of possible changes (i.e., by the number of consecutive pairs of layers in the multilayer

framework). The second definition is F � =1N

PNi=1 f �

i , where f �i is defined as the total number of modules of which node i

is a member at some point in the experiment session. We also indicate mean values and standard deviations.

25

Page 32: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Type of Flexibility Session Null Model t-statistic p-valueF 1 Connectional −10.62 1.2× 10−18

Nodal −47.56 1.1× 10−75

Temporal −0.64 0.512 Connectional −9.47 5.5× 10−16

Nodal −43.41 1.7× 10−71

Temporal −0.16 0.863 Connectional −9.86 1.1× 10−16

Nodal −46.95 4.5× 10−75

Temporal 0.77 0.43F � 1 Connectional −14.54 1.8× 10−27

Nodal −43.44 1.6× 10−71

Temporal 4.67 8.4× 10−6

2 Connectional −14.04 2.3× 10−26

Nodal −39.67 2.1× 10−67

Temporal 3.95 1.3× 10−4

3 Connectional −14.87 3.5× 10−28

Nodal −43.80 6.9× 10−72

Temporal 2.79 0.0061

Table 3: Results for one-sample t-tests on flexibility in cortical networks versus random network null modelsfor two different definitions of flexibility (F and F �). We report these results for the connectional, nodal, and temporal

null models. Negative t-statistics indicate that the flexibility of the null model network is greater than that of the cortical

network, and positive t-statistics indicate that the flexibility of the cortical network is greater than that of the null model

network.

26

Page 33: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Button Box Sequence

A

B

Trial Bin

Session 1 Session 2 Session 3

Mov

emen

t Tim

e (s

)

Figure 1: Experimental Setup and Learning (A) Schematic of the cued sequence production (CSP)task. The response or “button” box (left) had four response buttons that were color-coded to match thenotes on the “musical staff” (right) presented to the subject in the visual stimulus. This visual stimuluswas composed of 12 notes in sequence. Here we show one example of a single sequence. (B) Movementtime as a function of practiced trials, whose decreasing slope indicates that learning is occuring. (Wehave aggregated trials into 10 trial bins per session.)

27

Page 34: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Figure 2: Properties of the static and dynamic modularity indices Q± and Qml. The mean(column 1), standard deviation (column 2), and maximum (column 3) of the static modularity indexQ± is shown for (A) the large scale (entire experiment), (B) the mesoscopic scale (three experimentalsessions), and (C) the small scale (individual time windows) over the 100 samplings. Row (D) shows themean (column 1), standard deviation (column 2), and maximum (column 3) of the dynamic modularityindex Qml over the 100 samplings. In the figure, the standard deviation is abbreviated as STD. Boxplotsindicate 95% confidence intervals over subjects.

28

Page 35: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Figure 3: Effects of the coupling parameter C on the four diagnostics in this study: modularityindex Qml, number of modules n, module size (i.e., number of nodes) s, and module stationarity ζ. Wefirst averaged values over 100 ‘optimal’ partitions (see the discussion in the text), so this figure givesmean values of all diagnostics. The error bars indicate standard deviations over subjects and sessions.Colors indicate network type: cortical network (blue), temporal null model network (green), nodal nullmodel network (orange), and connectional null model network (red). Error bars for different networktypes at a given value of C (0.2, 0.4, 0.6, 0.8, 1, 1.2, 1.4, 1.6, 1.8, 2) are offset from each other for bettervisualization.

29

Page 36: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

Figure 4: Effect of the time window length T on the four diagnostics in this study: modularity indexQml, number of modules n, module size (i.e., number of nodes) s, and module stationarity ζ. We firstaveraged values over 100 ‘optimal’ partitions (see the discussion in the text), so this figure gives meanvalues of all diagnostics. The error bars indicate standard deviations over subjects and sessions. In thefigure, we give time windows in terms of the number of data points in the time series (i.e., the numberof TRs). Colors indicate network type: cortical network (blue), temporal null model network (green),nodal null model network (orange), and connectional null model network (red). Error bars for differentnetwork types at a given value of T (80, 90, 100, 110) are offset from each other for better visualization.

30

Page 37: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

References

[1] G. S. Snoddy, “Learning and stability: A psychophysical analysis of a case of motor learning with

clinical applications,” Journal of Applied Psychology, vol. 10, pp. 1–36, 1926.

[2] S. M. Smith, M. Jenkinson, M. W. Woolrich, C. F. Beckmann, T. E. J. Behrens, H. Johansen-Berg,

P. R. Bannister, M. De Luca, I. Drobnjak, D. E. Flitney, R. Niazy, J. Saunders, J. Vickers, Y. Zhang,

N. De Stefano, J. M. Brady, and P. M. Matthews, “Advances in functional and structural MR image

analysis and implementation as FSL,” Neuroimage, vol. 23, no. S1, pp. 208–219, 2004.

[3] D. S. Bassett and E. T. Bullmore, “Small-world brain networks,” Neuroscientist, vol. 12, pp. 512–523,

2006.

[4] D. S. Bassett and E. T. Bullmore, “Human brain networks in health and disease,” Curr Opin Neurol,

vol. 22, no. 4, pp. 340–347, 2009.

[5] E. Bullmore and O. Sporns, “Complex brain networks: Graph theoretical analysis of structural and

functional systems,” Nat Rev Neurosci, vol. 10, no. 3, pp. 186–198, 2009.

[6] J. Wang, L. Wang, Y. Zang, H. Yang, H. Tang, Q. Gong, Z. Chen, C. Zhu, and Y. He, “Parcellation-

dependent small-world brain functional networks: A resting-state fMRI study,” Hum Brain Mapp,

vol. 30, no. 5, pp. 1511–1523, 2009.

[7] A. Zalesky, A. Fornito, I. H. Harding, L. Cocchi, M. Yucel, C. Pantelis, and E. T. Bullmore, “Whole-

brain anatomical networks: Does the choice of nodes matter?,” Neuroimage, vol. 50, no. 3, pp. 970–

983, 2010.

[8] D. S. Bassett, J. A. Brown, V. Deshpande, J. M. Carlson, and S. T. Grafton, “Conserved and variable

architecture of human white matter connectivity,” Neuroimage, vol. 54, no. 2, pp. 1262–79, 2011.

[9] M. W. Woolrich, W. Jbabdi, B. Patenaude, M. Chappell, S. Makni, T. Behrens, C. Beckmann,

M. Jenkinson, and S. M. Smith, “Bayesian analysis of neuroimaging data in FSL,” Neuroimage,

vol. 45, pp. S173–S186, 2009.

[10] E. Bullmore, J. Fadili, M. Breakspear, R. Salvador, J. Suckling, and M. Brammer, “Wavelets and

statistical analysis of functional magnetic resonance images of the human brain,” Stat Methods Med

Res, vol. 12, no. 5, pp. 375–399, 2003.

31

Page 38: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

[11] E. Bullmore, J. Fadili, V. Maxim, L. Sendur, B. Whitcher, J. Suckling, M. Brammer, and M. Break-

spear, “Wavelets and functional magnetic resonance imaging of the human brain,” Neuroimage,

vol. 23, no. 1, pp. S234–S249, 2004.

[12] M. J. Brammer, “Multidimensional wavelet analysis of functional magnetic resonance images,” Hum

Brain Mapp, vol. 6, no. 5-6, pp. 378–382, 1998.

[13] S. Achard, R. Salvador, B. Whitcher, J. Suckling, and E. Bullmore, “A resilient, low-frequency,

small-world human brain functional network with highly connected association cortical hubs,” J

Neurosci, vol. 26, no. 1, pp. 63–72, 2006.

[14] D. S. Bassett, A. Meyer-Lindenberg, S. Achard, T. Duke, and E. Bullmore, “Adaptive reconfigu-

ration of fractal small-world human brain functional networks,” Proc Natl Acad Sci USA, vol. 103,

pp. 19518–19523, 2006.

[15] S. Achard and E. Bullmore, “Efficiency and cost of economical brain functional networks,” PLoS

Comput Biol, vol. 3, p. e17, 2007.

[16] S. Achard, D. S. Bassett, A. Meyer-Lindenberg, and E. Bullmore, “Fractal connectivity of long-

memory networks,” Phys Rev E, vol. 77, no. 3, p. 036104, 2008.

[17] D. S. Bassett, A. Meyer-Lindenberg, D. R. Weinberger, R. Coppola, and E. Bullmore, “Cogni-

tive fitness of cost-efficient brain functional networks,” Proc Natl Acad Sci USA, vol. 106, no. 28,

pp. 11747–11752, 2009.

[18] M. E. Lynall, D. S. Bassett, R. Kerwin, P. McKenna, U. Muller, and E. T. Bullmore, “Functional

connectivity and brain networks in schizophrenia,” J Neurosci, vol. 30, no. 28, pp. 9477–87, 2010.

[19] D. B. Percival and A. T. Walden, Wavelet Methods for Time Series Analysis. Cambridge University

Press, 2000.

[20] F. T. Sun, L. M. Miller, and M. D’Esposito, “Measuring interregional functional connectivity using

coherence and partial coherence analyses of fMRI data,” Neuroimage, vol. 21, no. 2, pp. 647–658,

2004.

[21] A. Barnes, E. T. Bullmore, and J. Suckling, “Endogenous human brain dynamics recover slowly

following cognitive effort,” PLoS One, vol. 4, no. 8, p. e6626, 2009.

32

Page 39: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

[22] M. J. Lowe, B. J. Mock, and J. A. Sorenson, “Functional connectivity in single and multislice

echoplanar imaging using resting state fluctuations,” Neuroimage, vol. 7, pp. 119–132, 1998.

[23] D. Cordes, V. M. Haughton, K. Arfanakis, G. J. Wendt, P. A. Turski, C. H. Moritz, M. A. Qugley,

and M. E. Meyerand, “Mapping functionally related regions of brain with functional connectivity

MR imaging,” Am J Neuroradiol, vol. 21, pp. 1636–1644, 2000.

[24] D. Cordes, V. M. Haughton, K. Arfanakis, J. D. Carew, P. A. Turski, C. H. Moritz, M. A. Quigley,

and M. E. Meyerand, “Frequencies contributing to functional connectivity in the cerebral cortex in

“resting-state” data,” Am J Neuroradiol, vol. 22, no. 7, pp. 1326–1333, 2001.

[25] D. A. Fair, B. L. Schlaggar, A. L. Cohen, F. M. Miezin, N. U. Dosenbach, K. K. Wenger, M. D. Fox,

A. Z. Snyder, M. E. Raichle, and S. E. Petersen, “A method for using blocked and event-related

fMRI data to study ‘resting state’ functional connectivity,” Neuroimage, vol. 35, no. 1, pp. 396–405,

2007.

[26] D. Meunier, S. Achard, A. Morcom, and E. Bullmore, “Age-related changes in modular organization

of human brain functional networks,” Neuroimage, vol. 44, no. 3, pp. 715–723, 2008.

[27] D. Meunier, R. Lambiotte, A. Fornito, K. D. Ersche, and E. T. Bullmore, “Hierarchical modularity

in human brain functional networks,” Front Neuroinformatics, vol. 3, p. 37, 2009.

[28] N. A. Rahman, A Course in Theoretical Statistics. Charles Griffin and Company, 1968.

[29] Y. He, Z. J. Chen, and A. C. Evans, “Small-world anatomical networks in the human brain revealed

by cortical thickness from MRI,” Cereb Cortex, vol. 17, pp. 2407–2419, 2007.

[30] Y. Benjamini and Y. Yekutieli, “The control of the false discovery rate in multiple testing under

dependency,” Ann Stat, vol. 29, pp. 1165–1188, 2001.

[31] C. R. Genovese, N. A. Lazar, and T. E. Nichols, “Thresholding of statistical maps in functional

neuroimaging using the false discovery rate,” Neuroimage, vol. 15, pp. 870–878, 2002.

[32] M. E. J. Newman, Networks: An Introduction. Oxford University Press, 2010.

[33] M. A. Porter, J.-P. Onnela, and P. J. Mucha, “Communities in networks,” Not Amer Math Soc,

vol. 56, no. 9, pp. 1082–1097, 1164–1166, 2009.

33

Page 40: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

[34] S. Fortunato, “Community detection in graphs,” Phys Rep, vol. 486, no. 3–5, pp. 75–174, 2010.

[35] B. H. Good, Y. A. de Montjoye, and A. Clauset, “Performance of modularity maximization in

practical contexts,” Phys Rev E, vol. 81, p. 046106, 2010.

[36] M. Girvan and M. E. Newman, “Community structure in social and biological networks,” Proc Natl

Acad Sci U S A, vol. 99, no. 12, pp. 7821–6, 2002.

[37] M. Girvan and M. E. Newman, “Community structure in social and biological networks,” Proc Natl

Acad Sci USA, vol. 99, no. 12, pp. 7821–6, 2002.

[38] U. Brandes, D. Delling, M. Gaertler, R. Gorke, M. Hoefer, Z. Nikoloski, and D. Wagner, “On

modularity clustering,” IEEE Transactions on Knowledge and Data Engineering, vol. 20, pp. 172–

188, 2008.

[39] Z. J. Chen, Y. He, P. Rosa-Neto, J. Germann, and A. C. Evans, “Revealing modular architecture

of human brain structural networks by using cortical thickness from MRI,” Cerebral Cortex, vol. 18,

no. 10, pp. 2374–2381, 2008.

[40] M. Rubinov and O. Sporns, “Complex network measures of brain connectivity: Uses and interpre-

tations,” Neuroimage, vol. 52, no. 3, pp. 1059–1069, 2009.

[41] V. D. Blondel, J. L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of community

hierarchies in large networks,” J Stat Mech, p. P10008, 2008.

[42] V. A. Traag and J. Bruggeman, “Community detection in networks with positive and negative links,”

Phys Rev E, vol. 80, no. 3 Pt 2, p. 036115, 2009.

[43] P. J. Mucha, T. Richardson, K. Macon, M. A. Porter, and J.-P. Onnela, “Community structure in

time-dependent, multiscale, and multiplex networks,” Science, vol. 328, no. 5980, pp. 876–878, 2010.

[44] D. S. Bassett, D. L. Greenfield, A. Meyer-Lindenberg, D. R. Weinberger, S. Moore, and E. Bullmore,

“Efficient physical embedding of topologically complex information processing networks in brains and

computer circuits,” PLoS Comput Biol, vol. 6, no. 4, p. e1000748, 2010.

[45] S. Maslov and K. Sneppen, “Specificity and stability in topology of protein networks,” Science,

vol. 296, no. 5569, pp. 910–913, 2002.

34

Page 41: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

[46] M. Molloy and B. Reed, “A critical point for random graphs with a given degree sequence,” Random

Structures and Algorithms, vol. 6, pp. 161–180, 1995.

[47] A. L. Traud, C. Frost, P. J. Mucha, and M. A. Porter, “Visualization of communities in networks,”

Chaos, vol. 19, p. 041104, 2009.

[48] T. M. J. Fruchtermann and E. M. Reingold, “Graph drawing by force-directed placement,” Software,

Practice and Experience, vol. 21, pp. 1129–1164, 1991.

[49] T. Kamada and S. Kawai, “An algorithm for drawing general undirected graphs,” Information

Processing Letters, vol. 31, pp. 7–15, 1988.

[50] K. J. Friston, E. Zarahn, O. Josephs, R. N. Henson, and A. M. Dale, “Stochastic designs in event-

related fMRI,” Neuroimage, vol. 10, no. 5, pp. 607–619, 1999.

[51] J. Benesty, J. Chen, and Y. Huang, “Estimation of the coherence function with the MVDR ap-

proach,” IEEE International Conference on Acoustic, Speech, and Signal Processing, 2006.

[52] G. Palla, A. Barabasi, and T. Vicsek, “Quantifying social group evolution,” Nature, vol. 446, pp. 664–

667, 2007.

[53] J. H. Callicott, V. S. Mattay, A. Bertolino, K. Finn, R. Coppola, J. A. Frank, T. E. Goldberg,

and D. R. Weinberger, “Physiological characteristics of capacity constraints in working memory as

revealed by functional MRI,” Cereb Cortex, vol. 9, no. 1, pp. 20–26, 1999.

[54] J. M. Lachin, “The role of measurement reliability in clinical trials,” Clin Trials, vol. 1, pp. 553–566,

2004.

[55] K. O. McGraw and S. P. Wong, “Forming inferences about some intraclass correlation coefficients,”

Psychological Methods, vol. 1, pp. 30–46, 1996.

[56] D. J. Fenn, M. A. Porter, M. McDonald, S. Williams, N. F. Johnson, and N. S. Jones, “Dynamic

communities in multichannel data: an application to the foreign exchange market during the 2007-

2008 credit crisis,” Chaos, vol. 19, no. 3, p. 033119, 2009.

[57] S. Fortunato and M. Barthelemy, “Resolution limit in community detection,” Proc Natl Acad Sci

USA, vol. 104, no. 1, pp. 36–41, 2007.

35

Page 42: Dynamic reconfiguration of human brain networks during learningmason/papers/brains_final.pdf · 2011-05-03 · Dynamic reconfiguration of human brain networks during learning Danielle

[58] A. Zalesky, A. Fornito, and E. T. Bullmore, “Network-based statistic: Identifying differences in brain

networks,” Neuroimage, vol. 53, no. 4, pp. 1197–207, 2010.

[59] J. Richiardi, H. Eryilmaz, S. Schwartz, P. Vuilleumier, and D. Van De Ville, “Decoding brain states

from fMRI connectivity graphs,” Neuroimage, vol. doi:10.1016/j.neuroimage.2010.05.081, 2010.

[60] S. V. Shinkareva, V. Gudkov, and J. Wang, “A network analysis approach to fMRI condition-specific

functional connectivity,” arXiv:1008.0590, 2010.

[61] B. W. Kernighan and S. Lin, “An efficient heuristic procedure for partitioning graphs,” The Bell

System Technical Journal, vol. 49, pp. 291–307, 1970.

[62] T. Richardson, P. J. Mucha, and M. A. Porter, “Spectral tripartitioning of networks,” Physical

Review E, vol. 80, no. 3, p. 036111, 2009.

[63] M. E. J. Newman, “Finding community structure in networks using the eigenvectors of matrices,”

Physical Review E, vol. 74, p. 036104, 2006.

Addendum

We thank two anonymous reviewers for helpful comments on this manuscript. This work was supported by

the David and Lucile Packard Foundation, PHS Grant NS44393, the Institute for Collaborative Biotech-

nologies through contract no. W911NF-09-D-0001 from the U.S. Army Research Office, and the NSF

(DMS-0645369). M.A.P. acknowledges a research award (#220020177) from the James S. McDonnell

Foundation.

We thank Aaron Clauset for useful discussions and John Bushnell for technical support.

Competing Interests: The authors declare that they have no competing financial interests.

Correspondence: Correspondence and requests for materials should be addressed to D.S.B. (email: dbas-

[email protected]).

36


Recommended